# Benchmarks

> Grit vs Git on realistic workloads, from the committed grit-bench baseline.

Performance work is measured on the factory VM and in CI against the **system** `git` binary. Micro-benchmarks live in `grit-lib` (Criterion) and command-level scenarios in `grit-bench` (`grit-utils`).

For a map of **grit-lib** modules, **grit** commands, common workflows, and every **`grit-bench daily`** scenario id, see [`grit-utils/MATRIX.md`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/MATRIX.md). The same ids appear with **time and memory ratios** on the [Performance overview (daily matrix)](https://grit-scm.com/docs/benchmarks/performance-overview.md/index.md) page.

Command-level numbers below come from the committed **grit-bench** JSON baselines listed in `content/docs/site.toml`. Scenario suites use **hyperfine** against the system **`git`** binary and a release **`grit`** build on the same machine; individual studies below document their sampling method. Times are wall-clock milliseconds. The daily census behind the [Performance overview](https://grit-scm.com/docs/benchmarks/performance-overview.md/index.md) measures the shipped **`dist`** build since 2026-10-28; see [Benchmarking](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#grit-bench-daily-factory-census) for what that switch did to every baseline.

## Push reads only commit headers while finding what the remote has (2026-11-03)

`push/refspec@git-git` pushes 50 new commits to a new branch over smart HTTP, against a copy of git.git served by `git http-backend`. The harness dates those commits in 2001, before any of git.git's history. Before it builds the pack, a push walks the remote's tips and the new commits in commit-date order, stopping once no queued commit is reachable only from the new ones. Here that order puts the new commits last, so the walk reads all of the remote's history first: about 75,000 commits in a client with no commit-graph. Git makes the same walk, and spends about 510 ms of its 1.05 s on it; most of the rest is the server.

Grit read each of those commits through the general object read: it set up a pack reader and checked the pack directory for every object, then inflated the whole commit, message included. The walk needs only the tree, parents and committer date. It now reads just the commit's header block from the pack, and falls back to a whole read for loose or deltified commits. The header-block reader also stops inflating at the blank line that ends the headers, stepping through the compressed input 256 bytes at a time. Before, it inflated up to 1 KiB plus slack for every commit, which for most commits is the whole message. fsck's connectivity pass and merge-base's tip seeding use the same reader.

Rows from `grit-bench daily` on `release` builds against the system `git`, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0, 10 runs per row. "before" is `3517a637f`, the build fix this change sits on. The before and after reports were taken back to back, and Git's own times differ by up to 3% between them. Values are **milliseconds / MB**. Reports: `daily-push-refspec-6936-before.json` → `daily-push-refspec-6936-after.json`.

| Scenario | git ms / MB | before ms / MB | after ms / MB | time vs git |
| --- | ---: | ---: | ---: | ---: |
| `push/refspec@git-git` | 1031.7 / 100.3 | 1132.7 / 77.7 | **989.1** / 77.8 | 1.10 → 0.94 |
| `push/force-with-lease@git-git` | 2114.8 / 334.2 | 1127.7 / 75.1 | **984.4** / 74.7 | 0.53 → 0.45 |
| `push/incremental-http@git-git` | 1616.2 / 334.1 | 1227.4 / 91.0 | **1094.0** / 91.5 | 0.76 → 0.66 |
| `push/incremental@git-git` | 2135.6 / 334.3 | 1712.2 / 90.4 | **1598.2** / 91.0 | 0.80 → 0.73 |
| `fsck/full@git-git` | 27492.5 / 444.2 | 13292.2 / 593.0 | 13577.4 / 595.1 | 0.48 → 0.49 |
| `merge-base/branches@git-git` | 15.2 / 19.7 | 4.1 / 19.3 | 4.2 / 19.5 | 0.27 → 0.27 |
| `log/json-n10-loose-tags@git-git` | 12.1 / 17.4 | 13.8 / 21.4 | 14.0 / 21.3 | 1.14 → 1.17 |
| `log/json-n10-packed-tags@git-git` | 10.0 / 17.4 | 2.8 / 13.6 | 2.8 / 13.8 | 0.28 → 0.28 |
| `revs/count@git-git` | 497.6 / 62.9 | 711.8 / 122.3 | 744.1 / 122.6 | 1.43 → 1.46 |

The first four rows are pushes. Interleaved hyperfine runs show no change in the others: `grit fsck` took 13.13 and 13.32 s before and 13.17 and 12.96 s after, and `grit revs --count HEAD` 717 and 724 ms before and 729 and 708 ms after. With a hand-run `git http-backend` front end and 10 hyperfine runs per command, the target push went from 1.287 s to 0.988 s against Git's 1.054 s, and grit's own CPU time from 625 to 400 ms. Peak memory did not move.

```bash
./target/release/grit-bench daily --only push/refspec@git-git,push/force-with-lease@git-git,push/incremental-http@git-git,push/incremental@git-git,fsck/full@git-git,merge-base/branches@git-git,log/json-n10-loose-tags@git-git,log/json-n10-packed-tags@git-git,revs/count@git-git \
  --min-runs 10 --grit ./target/release/grit --format json --output grit-utils/baselines/daily-push-refspec-6936-after.json
```

## Writing the commit-graph reads each commit once (2026-11-03)

`repack/commit-graph@git-git` times `grit gc commit-graph` against `git commit-graph write --reachable --changed-paths` on git.git (75,031 commits, one pack) with a changed-path graph already in place, so both tools reuse every Bloom filter. Grit took twice Git's time, and almost all of it went to zlib: the writer walked the refs to collect the reachable commits, then read every commit a second time to load its tree, parents and date, reopening the single-pack reader for each read.

The walk now keeps each commit's graph data as it goes, reading every object once through one pack reader for the whole walk. Each commit is still read whole, zlib checksum included, and parsed from the object database rather than from the graph it replaces, so a damaged commit fails the write as it does in Git, and rewriting a damaged graph repairs it. (Inflating only the commit headers would be faster still, but would accept a commit whose message is corrupt.) A commit that exists but cannot be read used to be left out of the graph silently; it is now an error.

When filters have to be computed (no graph yet, or `--max-new-filters` left some out), the writer used to diff each commit against its first parent in object-id order, which visits history at random, and read the parent commit again to find its tree. It now diffs in generation order, so consecutive diffs share most of their trees and find them in the object database's tree cache, and takes the parent's tree from the walk. The budget still goes to the first commits in object-id order, and the file is byte-for-byte what the previous writer produced. On the same git.git clone with no graph, `grit gc commit-graph` drops from 11.8 to 5.7 s (Git: 4.6 s in the same sitting), at 380 MB peak RSS instead of 361 MB (Git: 366 MB).

Rows from `grit-bench daily` on `release` builds against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0. "base" is `8cb4162ac`; "after" is the branch's code at `e5c51b48c`. Times are medians of 10 runs; values are **milliseconds / MB**. The git column is from the second report, and each ratio uses the git time of its own report. Reports: `daily-commit-graph-6930-before.json` → `daily-commit-graph-6930-after.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | time vs git |
| --- | ---: | ---: | ---: | ---: |
| `repack/commit-graph@git-git` | 619.88 / 75.4 | 1244.10 / 107.1 | **493.35** / 106.3 | 2.01 → 0.80 |
| `repack/full@git-git` | 4076.84 / 481.0 | 6334.22 / 565.1 | 6357.35 / 564.3 | 1.58 → 1.56 |
| `gc/default@synth-10k` | 326.92 / 49.3 | 286.47 / 53.6 | 266.33 / 53.9 | 0.85 → 0.81 |

The `gc/default@synth-10k` cells are medians of the runs that repacked (9 of 10 before, 4 of 10 after). In the others the prepare step had not seeded loose objects and `grit gc` only rewrote the 1,000-commit graph, in 19 ms before and 11 ms after; the reports' own medians mix both kinds of run.

```bash
GRIT_BENCH_REPORT_GRIT_COMMIT=<measured revision> ./target/release/grit-bench daily --offline --only repack/commit-graph@git-git,repack/full@git-git,gc/default@synth-10k \
  --min-runs 10 --grit ./target/release/grit --format json --output grit-utils/baselines/daily-commit-graph-6930-after.json
```

## Commit counts read only parent links (2026-11-03)

`revs/count@git-git` runs `grit revs --count HEAD` on git.git (v2.47.0, 75,031 commits in one pack, no commit-graph or bitmap) against `git rev-list --count HEAD`. Grit took 701 ms and peaked at 121 MB, against Git's 498 ms and 63 MB. Two costs made up the gap.

- **Every object read stat'ed the pack directory.** The first time a pack's bytes were mapped, the pack cache treated it like a replaced pack and dropped the thread's trusted pack listing. Nothing restores that trust until the store's epoch changes, so each later read checked the `objects/pack` mtime (75,000 `statx` calls) and reopened the single-pack reader through its path-keyed caches instead of its memo. A first mapping now leaves the listing, the parsed index and the pack hint alone. Only a pack whose bytes were already cached is invalidated. Every command that reads many packed objects gains from this: `log/path`, `revs/count-path` and `revs/count-author` below.
- **The count went through the general walk.** That walk parsed and kept every commit in full (date-order queue, owned commit data, message strings) for an answer that needs only parent links. An unfiltered count (no excludes, limits, ordering, path or commit filters, and no commit-graph or bitmap to answer it) now walks a visited set over parent lists. A commit the single local pack stores whole is inflated into one reused buffer, and only its `tree` and `parent` headers are parsed. Deltified and loose commits go through the object database. Grafts and shallow boundaries apply as in the filtered count. Replacement refs fall back to the full walk. When a commit-graph exists but a tip is newer than it, the graph-only count declines, so the walk takes parents from the graph for every commit the graph lists and reads only the newer ones. On git.git with a reachable commit-graph and one commit on top (`hyperfine`, release builds, 15 runs), `grit revs --count HEAD` takes 29 ms, against 101 ms before and 45 ms for Git.

Each entry is inflated to the end of its zlib stream, so a commit whose stream is corrupt past its headers fails the count as it fails Git's. Inflating only the `tree` and two `parent` lines was about 65 ms faster here (296 ms against 359 ms on release builds), but it accepted such a commit without checking its checksum, so the count pays for the full inflate. What remains is mostly zlib.

Rows from `grit-bench daily` on `dist` builds of `grit-cli` with `grit-utils`, against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0. "base" is `8cb4162ac` with the stash build fix from `ab28c10a1`. Times are medians of all samples from two rounds of 10 runs (base, after, after, base), and peak RSS is the median of the rounds' peaks. Values are **milliseconds / MB**. Reports (first round): `daily-revs-count-6954-before.json` → `daily-revs-count-6954-after.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | time vs git | RSS vs git |
| --- | ---: | ---: | ---: | ---: | ---: |
| `revs/count@git-git` | 497.8 / 62.9 | 701.3 / 121.2 | **361.4** / **55.9** | 1.41 → 0.73 | 1.93× → 0.89× |
| `revs/count-path@git-git` | 386.0 / 250.8 | 445.2 / 274.6 | 367.8 / 274.8 | 1.15 → 0.95 | 1.09× → 1.10× |
| `revs/count-author@deep-history` | 109.7 / 21.8 | 189.2 / 22.9 | 117.3 / 23.1 | 1.73 → 1.07 | 1.05× → 1.06× |
| `revs/count-author@deep-history-cg` | 115.6 / 25.0 | 190.2 / 23.0 | 118.4 / 23.1 | 1.64 → 1.02 | 0.92× → 0.93× |
| `revs/count-range-commit-graph@git-git` | 2.43 / 19.5 | 2.61 / 14.2 | 2.64 / 14.6 | 1.07 → 1.09 | 0.73× → 0.74× |
| `revs/count-since@deep-history` | 1.47 / 8.3 | 2.15 / 10.0 | 2.06 / 10.4 | 1.46 → 1.40 | 1.20× → 1.25× |
| `revs/author-n50@deep-history` | 1.62 / 9.0 | 2.72 / 12.7 | 2.61 / 12.2 | 1.68 → 1.61 | 1.42× → 1.36× |
| `revs/author-n50@deep-history-cg` | 1.89 / 12.3 | 2.90 / 15.8 | 2.82 / 15.5 | 1.53 → 1.49 | 1.29× → 1.26× |
| `revs/left-right-count@deep-history` | 2.30 / 9.0 | 4.46 / 11.6 | 3.59 / 11.4 | 1.94 → 1.56 | 1.29× → 1.27× |
| `log/all@git-git` | 44.5 / 20.9 | 58.3 / 26.8 | 43.9 / 27.0 | 1.31 → 0.99 | 1.28× → 1.29× |
| `log/path@git-git` | 320.3 / 260.8 | 397.3 / 276.7 | 324.5 / 276.9 | 1.24 → 1.01 | 1.06× → 1.06× |
| `log/author@git-git` | 2.34 / 11.1 | 3.95 / 16.5 | 3.82 / 15.9 | 1.69 → 1.63 | 1.48× → 1.42× |
| `log/oneline-100@git-git` | 2.88 / 15.0 | 3.96 / 17.0 | 3.59 / 16.9 | 1.37 → 1.24 | 1.14× → 1.13× |
| `blame/file@git-git` | 271.7 / 243.9 | 225.0 / 265.0 | 220.7 / 265.2 | 0.83 → 0.81 | 1.09× → 1.09× |
| `merge-base/branches@git-git` | 15.2 / 19.8 | 3.89 / 17.9 | 3.45 / 17.3 | 0.26 → 0.23 | 0.90× → 0.87× |
| `show/head@git-git` | 1.80 / 9.0 | 2.34 / 10.4 | 2.22 / 10.0 | 1.30 → 1.24 | 1.15× → 1.11× |
| `cat/blob@git-git` | 1.36 / 7.2 | 1.93 / 9.4 | 1.83 / 9.5 | 1.42 → 1.34 | 1.30× → 1.31× |

The rows under 5 ms moved within noise (±0.1 ms between rounds).

```bash
./target/dist/grit-bench daily --only revs/count@git-git,revs/count-path@git-git,revs/count-author@deep-history,revs/count-author@deep-history-cg,revs/count-range-commit-graph@git-git,revs/count-since@deep-history,revs/author-n50@deep-history,revs/author-n50@deep-history-cg,revs/left-right-count@deep-history,log/all@git-git,log/author@git-git,log/oneline-100@git-git,log/path@git-git,blame/file@git-git,cat/blob@git-git,merge-base/branches@git-git,show/head@git-git \
  --min-runs 10 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-revs-count-6954-after.json
```

## Repacking skips tree entries earlier versions repeat (2026-11-03)

`repack/full@git-git` runs `grit gc repack` against `git repack -ad` on a git.git worktree with one pack, no bitmap and no commit-graph, so every object is found by walking history. The census had grit at 6.3 s against Git's 4.6 s. The object walk was single-threaded and took most of the time. Its 147,000 trees hold 2.2 GB of entries, and the root tree alone (about 535 entries) is expanded some 75,000 times. Almost all of those entries name objects the walk has already met.

Without a filter, every entry of a tree the walk has expanded names an object that is visited from then on. The walk now keeps the last tree it expanded at each path, keyed by the path's name hash, and compares the next version there against it. Entries in the bytes both versions share at the start and the end are neither parsed nor looked up. Between those two ranges, a cursor steps through the earlier version in name order, and any entry it repeats byte for byte is skipped. Entry offsets carry over from the earlier version, so a skipped range is always whole entries. Which version is compared only decides how much is skipped, never what is listed. Entries are located by a small span parser that builds no object id for skipped entries and leaves anything unusual to `TreeRefIter`, with the same result and errors. The pack is byte-identical, and a randomized test over 300 edited tree versions and a `git rev-list --objects --all` round trip check the object list.

Rows from one `grit-bench daily` sitting on `release` builds, against the system `git` 2.43.0 on a four-vCPU Intel Xeon VM. "base" is the tree read-ahead change from [the bundling section](#bundling-every-ref-without-a-bitmap-2026-11-03) that this builds on. On `main` in the same sitting, `repack/full@git-git` took 6249 ms / 565 MB. Times are medians of 10 runs (target rows) or 5 runs (the others) after one warm-up. Values are **milliseconds / MB**. Reports: `daily-repack-full-6961-before.json` → `daily-repack-full-6961-after.json` and `daily-repack-full-6961-related-before.json` → `daily-repack-full-6961-related-after.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | after vs git |
| --- | ---: | ---: | ---: | ---: |
| `repack/full@git-git` | 4484.3 / 475 | 3672.0 / 579 | **2374.2** / 569 | 0.53 |
| `repack/commit-graph@git-git` | 635.6 / 109 | 1276.6 / 141 | 1253.4 / 141 | 1.97 |
| `bundle/create-all@git-git` | 4347.5 / 563 | 3849.2 / 725 | **2616.2** / 711 | 0.60 |
| `bundle/create@synth-10k` | 253.5 / 44 | 123.3 / 58 | 121.3 / 59 | 0.48 |
| `gc/default@synth-10k` | 325.1 / 49 | 137.0 / 60 | 139.9 / 57 | 0.43 |
| `clone/full@git-git` | 14690.6 / 125 | 7710.5 / 154 | 7783.2 / 155 | 0.53 |
| `clone/blobless@git-git` | 8588.5 / 73 | 7005.7 / 71 | 7134.3 / 71 | 0.83 |
| `clone/shallow@git-git` | 1570.7 / 40 | 1607.4 / 57 | 1607.2 / 58 | 1.02 |
| `clone/bare@synth-10k` | 825.9 / 37 | 692.9 / 29 | 692.0 / 29 | 0.84 |
| `clone/local@synth-10k` | 945.2 / 37 | 846.8 / 34 | 847.7 / 33 | 0.90 |
| `fetch/incremental@git-git` | 104.3 / 71 | 55.0 / 49 | 55.6 / 49 | 0.53 |
| `fetch/one-branch@git-git` | 103.7 / 71 | 55.3 / 50 | 55.4 / 49 | 0.53 |
| `push/incremental@git-git` | 2117.0 / 334 | 1693.9 / 91 | 1710.4 / 91 | 0.81 |
| `push/refspec@git-git` | 1064.0 / 100 | 1151.6 / 78 | 1163.5 / 78 | 1.09 |
| `fsck/full@git-git` | 25901.7 / 669 | 12053.1 / 638 | 12379.4 / 638 | 0.48 |
| `revs/count@git-git` | 500.4 / 97 | 696.6 / 156 | 711.7 / 156 | 1.42 |
| `log/path@git-git` | 313.4 / 198 | 379.8 / 217 | 384.3 / 216 | 1.23 |
| `diff/worktree@deep-tree` | 97.7 / 41 | 67.5 / 41 | 70.6 / 41 | 0.72 |
| `status/clean@synth-100k` | 56.8 / 30 | 44.6 / 33 | 44.4 / 33 | 0.78 |

Repack and bundling every ref gain about 1.3 s each. The other rows move within noise: they send few objects, filter the walk (a blobless clone, where the skipping is off), spend their time receiving and checking out the pack, or do not walk objects at all.

```bash
./target/release/grit-bench daily --only repack/full@git-git,repack/commit-graph@git-git \
  --grit ./target/release/grit --warmup 1 --min-runs 10 --output grit-utils/baselines/daily-repack-full-6961-after.json
```

## Bundling every ref without a bitmap (2026-11-03)

`bundle/create-all@git-git` runs `grit bundle create <file> --all` on a fully packed git.git clone that has no reachability bitmap and no commit-graph, so every object has to be found by walking history. The daily census had grit at 7.8 s against Git's 4.4 s. Most of that cost was in the object walk and delta planning that bundles share with fetch serving, push and repack, so this change made that shared path faster:

- **No needless commit walks.** `--all` with no exclusions has no boundary commits, but grit still ran a full `rev-list --boundary` over all 75,000 commits before building the pack. The have-side pre-walk that bounds a push also read every commit when there were no haves at all, only to return nothing. Both are now skipped.
- **Cheaper tree entries.** The object walk now carries each directory's pack name hash instead of building a path string for every subtree and new blob. It skips tree entries it has already visited, reads trees without a separate kind lookup, and parses entry modes without UTF-8 validation.
- **Parallel tree reads.** One pack reader is held for the whole walk instead of being reopened per object, and objects are no longer copied into the session read cache. Worker threads (`pack.threads` minus the walk's own) inflate the next runs of queued trees while the walk parses the current ones. Results are consumed in queue order, so the bundle is byte-for-byte the same as with one thread.
- **Linear delta planning.** Capping delta chains at `pack.depth` used to walk the whole chain from every tip. It now visits each object once, and a randomized test checks it against the old walk. Reusable on-disk deltas are scanned in pack order, which takes each entry's end and CRC from the offset table instead of searching for them.

The object walk took about 6 s of the base's 9.2 s and now takes 2.8 s. Delta planning takes 1.0 s instead of 1.4 s. Peak live heap drops from about 300 MB to 200 MB. Peak RSS is unchanged at about 720 MB against Git's 530 MB. That memory is the mapped pack (about 260 MB) plus allocator retention: with `pack.threads=1`, both builds still peak at 660 to 690 MB.

Rows from one `grit-bench daily` sitting on two `release` builds (base `aa19aba66`, after `05349d540`), against the system `git` 2.43.0 on a four-vCPU Intel Xeon VM. Times are medians of 5 runs after one warm-up. Values are **milliseconds / MB**. Reports: `daily-bundle-create-all-6895-before.json` → `daily-bundle-create-all-6895-after.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | after vs git |
| --- | ---: | ---: | ---: | ---: |
| `bundle/create-all@git-git` | 4387.9 / 529 | 9152.4 / 719 | **3929.5** / 724 | 0.90 |
| `bundle/create@synth-10k` | 255.9 / 44 | 279.1 / 55 | **131.4** / 58 | 0.51 |
| `bundle/verify@synth-10k` | 1.4 / 7 | 2.7 / 9 | 2.6 / 9 | 1.88 |
| `repack/full@git-git` | 4509.5 / 475 | 6650.1 / 565 | **3871.2** / 573 | 0.86 |
| `gc/default@synth-10k` | 335.2 / 49 | 288.2 / 55 | **142.4** / 58 | 0.42 |
| `clone/full@git-git` | 14775.2 / 120 | 8075.3 / 156 | 7914.3 / 157 | 0.54 |
| `clone/blobless@git-git` | 8715.0 / 75 | 7377.3 / 71 | 7306.2 / 71 | 0.84 |
| `clone/shallow@git-git` | 1602.6 / 40 | 1631.5 / 55 | 1631.7 / 57 | 1.02 |
| `clone/bare@synth-10k` | 827.5 / 37 | 701.4 / 29 | 694.1 / 29 | 0.84 |
| `clone/local@synth-10k` | 974.0 / 37 | 858.7 / 32 | 860.1 / 33 | 0.88 |
| `clone/http@synth-10k` | 598.8 / 37 | 469.7 / 33 | 460.8 / 34 | 0.77 |
| `fetch/incremental@git-git` | 103.8 / 71 | 57.7 / 49 | 56.4 / 50 | 0.54 |
| `fetch/deepen@git-git` | 692.9 / 272 | 175.8 / 58 | 180.6 / 59 | 0.26 |
| `push/incremental@git-git` | 2157.5 / 322 | 1754.1 / 91 | 1730.7 / 91 | 0.80 |
| `push/refspec@git-git` | 1072.9 / 101 | 1176.2 / 78 | 1172.9 / 78 | 1.09 |

The bundle, repack and gc rows gain from the shared path. The other rows move within noise, because their time goes elsewhere (receiving, indexing and checking out the pack) or they send few objects.

```bash
S=bundle/create-all@git-git,bundle/create@synth-10k,bundle/verify@synth-10k,clone/bare@synth-10k,clone/local@synth-10k,clone/http@synth-10k,clone/full@git-git,clone/blobless@git-git,clone/shallow@git-git,fetch/incremental@git-git,fetch/deepen@git-git,push/incremental@git-git,push/refspec@git-git,repack/full@git-git,gc/default@synth-10k
./target/release/grit-bench daily --only "$S" --grit ./target/release/grit --http-server ./target/release/grit-http-server \
  --warmup 1 --min-runs 5 --output grit-utils/baselines/daily-bundle-create-all-6895-after.json
```

## Pick, revert and merge join the hot text layout (2026-11-02)

`pick/one@synth-10k` cherry-picks a topic commit that rewrites 2,000 of 10,000 files onto `main`. Grit already took under half Git's time, but peaked at 18.8 MB on a `dist` build and 22.3 MB on `release`, against Git's 9.4 MB. Little of that was data: the heap peaks at 4.7 MB. On `release`, 12.5 MB of the binary was still resident when `grit pick` exited (8.6 MB on `dist`).

The [hot text layout](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#hot-text-layout) had never profiled a pick. Its workloads covered status, diff, log, refs, staging, committing, switching and stashing. The code that only replays and merges commits (the replay sequence, the tree merge, the work-tree guard and the parallel checkout writes) therefore stayed scattered over plain `.text`, and each executed function made its 64 KB fault-around window resident. With fault-around temporarily set to one page, the `release` pick peaks at 12.1 MB instead of 20.7 MB, and Git at 6.0 MB instead of 8.9 MB. Most of the gap was fault-around over scattered code.

`scripts/gen_hot_text_layout.py` now profiles three more workloads: `grit pick origin/topic`, `grit revert HEAD` and `grit merge origin/topic`. They run on a topic branch, forked before the last commit on `main`, that edits the first line of every fifth file. The new workloads come last in the workload order, so the Gray code order and the layout search split the earlier workloads' runs as little as possible. Two details decide whether the lists match what the census runs:

- **The untracked cache.** With `core.untrackedCache` unset, grit adds an untracked cache to the index of a repository with 4,096 or more entries. The census's pick and merge fixtures leave the setting unset, so every pick there parses and writes that index extension. The generator's other workloads turn the cache off, as the census's status fixtures do. The new workloads run with the setting unset and an untracked cache already in the index.
- **Feature sets.** Building `grit` together with `grit-utils`, as the census does, turns on grit-lib's `serde-models` feature, and its extra impls shift the impl disambiguators (`Ms<n>_`) in symbol names, which the patterns do not wildcard. Shipped binaries build `grit-cli` alone. The generator now profiles both builds of both profiles.

On `dist`, the library section grows from 3.6 to 4.0 MB and the CLI section from 2.2 to 2.3 MB. A section only costs memory where a command runs its code. After the change, the pick's resident binary pages at exit drop to 6.5 MB on `release` and 5.6 MB on `dist`.

Rows from `grit-bench daily` on `dist` builds of `grit-cli` with `grit-utils` (the census build), against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0. "base" is `85f9bca2b` with the `rebase_run_request` import fix from `67b7ad8df`. Times are medians of all samples from two alternating rounds of 15 runs, and peak RSS is the median of the rounds' peaks. Values are **milliseconds / MB**. Reports (first round): `daily-pick-one-6676-before.json` → `daily-pick-one-6676-after.json` and `daily-pick-one-6676-related-before.json` → `daily-pick-one-6676-related-after.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | time vs git | RSS vs git |
| --- | ---: | ---: | ---: | ---: | ---: |
| `pick/one@synth-10k` | 117.01 / 9.4 | 54.25 / 18.8 | **53.66** / **16.3** | 0.46 → 0.46 | 2.00× → 1.74× |
| `pick/one@synth-100k` | 192.72 / 38.9 | 130.82 / 34.8 | 130.75 / 31.7 | 0.68 → 0.68 | 0.89× → 0.81× |
| `revert/one@synth-10k` | 111.24 / 9.4 | 56.12 / 20.1 | 56.32 / 17.6 | 0.50 → 0.51 | 2.13× → 1.87× |
| `merge/clean@synth-10k` | 104.32 / 14.2 | 53.13 / 18.3 | 52.52 / 15.6 | 0.51 → 0.50 | 1.29× → 1.10× |
| `merge/clean@synth-100k` | 289.65 / 46.3 | 144.09 / 41.7 | 141.26 / 39.4 | 0.50 → 0.49 | 0.90× → 0.85× |
| `merge/content-both-sides@wide-20k` | 1332.62 / 66.3 | 1254.90 / 77.5 | 1253.15 / 74.1 | 0.94 → 0.94 | 1.17× → 1.12× |
| `rebase/onto@synth-10k` | 50.31 / 14.2 | 11.55 / 12.8 | 12.00 / 12.6 | 0.23 → 0.24 | 0.90× → 0.89× |
| `history/reword@synth-10k` | 6352.79 / 30.6 | 137.65 / 19.6 | 138.05 / 19.1 | 0.02 → 0.02 | 0.64× → 0.62× |
| `status/clean@synth-10k` | 6.97 / 9.6 | 7.71 / 12.9 | 7.19 / 11.3 | 1.11 → 1.03 | 1.35× → 1.18× |
| `status/dirty@synth-10k` | 10.51 / 9.7 | 8.11 / 13.2 | 7.63 / 11.4 | 0.77 → 0.73 | 1.36× → 1.18× |
| `status/renamed@synth-10k` | 7.25 / 9.7 | 7.70 / 12.2 | 7.69 / 13.7 | 1.06 → 1.06 | 1.26× → 1.41× |
| `status/clean@small` | 1.95 / 7.5 | 2.86 / 9.7 | 2.79 / 9.6 | 1.47 → 1.43 | 1.30× → 1.28× |
| `diff/worktree@synth-10k` | 7.06 / 9.9 | 6.83 / 12.5 | 6.94 / 12.6 | 0.97 → 0.98 | 1.26× → 1.27× |
| `show/head@synth-10k` | 2.02 / 8.9 | 2.51 / 10.3 | 2.46 / 10.5 | 1.24 → 1.22 | 1.15× → 1.18× |
| `log/oneline-100@synth-100k` | 2.26 / 9.8 | 4.43 / 12.5 | 4.35 / 12.4 | 1.96 → 1.92 | 1.28× → 1.27× |
| `branch/list@many-refs` | 1.32 / 6.7 | 1.92 / 9.0 | 1.86 / 9.1 | 1.46 → 1.42 | 1.35× → 1.36× |
| `branch/merged@many-refs` | 1.49 / 7.7 | 1.96 / 9.7 | 1.96 / 9.8 | 1.31 → 1.31 | 1.26× → 1.28× |
| `branch/create@many-refs` | 1.53 / 7.8 | 2.35 / 10.0 | 2.36 / 9.4 | 1.54 → 1.55 | 1.29× → 1.21× |
| `config/get@small` | 1.14 / 6.7 | 1.59 / 7.8 | 1.55 / 7.9 | 1.39 → 1.36 | 1.16× → 1.18× |
| `add/modified@synth-10k` | 16.75 / 12.4 | 9.98 / 17.4 | 9.95 / 17.3 | 0.60 → 0.59 | 1.40× → 1.39× |
| `restore/paths@synth-10k` | 12.13 / 12.6 | 13.52 / 16.4 | 13.66 / 16.2 | 1.11 → 1.13 | 1.30× → 1.29× |
| `commit/small@synth-10k` | 27.16 / 13.9 | 16.81 / 14.7 | 16.31 / 17.3 | 0.62 → 0.60 | 1.06× → 1.24× |
| `reset/hard@synth-10k` | 15.93 / 11.2 | 14.78 / 15.7 | 14.74 / 15.8 | 0.93 → 0.93 | 1.41× → 1.41× |
| `switch/branch@synth-10k` | 21.94 / 10.7 | 16.04 / 13.2 | 15.34 / 12.9 | 0.73 → 0.70 | 1.24× → 1.21× |
| `switch/branch@synth-100k` | 154.52 / 37.6 | 84.75 / 31.9 | 81.96 / 31.8 | 0.55 → 0.53 | 0.85× → 0.85× |
| `stash/push-pop@synth-10k` | 206.02 / 18.3 | 129.76 / 17.8 | 128.83 / 17.9 | 0.63 → 0.63 | 0.97× → 0.97× |

Run time moved within noise. The `commit/small@synth-10k` and `status/renamed@synth-10k` peaks are bimodal for either build, as described in [the status section below](#status-code-layout-by-workload-set-2026-11-01). Measured directly, the layout lowers both. With the harness's own prepare step and 8 interleaved pairs, the commit's resident binary pages at exit drop from 6.37 to 5.93 MB, with an identical heap. Over 12 interleaved pairs of `grit status` after `git reset --hard`, they drop from 5.46 to 5.31 MB, and the peak from 14.2 to 14.1 MB.

`status/clean@synth-10k` peaks are bimodal the same way. A row's peak is the largest of its runs, and the high mode of about 13 MB appears exactly when one or more of the first timed runs take 16 to 30 ms instead of 7 ms. Those runs rehash racy index entries right after the harness has reset the fixture. Over 16 alternating single-row rounds on the rebased tree (`6083ca2bb`), in both orders, the base hit the high mode in 4 and the change in 6, each time with slow runs. In every round without them, the base peaked at 10.85 to 11.08 MiB and the change at 10.67 to 10.81 MiB. Measured directly after the same prepare step (30 interleaved pairs), the change keeps 4,120 KiB of the binary resident instead of 4,316 KiB. Forcing the racy rehash on every run (15 pairs), the peak is 11,420 KiB instead of 11,508 KiB, with 4,264 KiB of binary pages instead of 4,392 KiB and an identical heap. The `release` build and the shipped `dist` build show the same direction on both paths.

The other builds show the same change. On `release` (one round of 15), `pick/one@synth-10k` peaks at 17.3 MB instead of 22.3 MB, `revert/one@synth-10k` at 18.6 MB instead of 23.9 MB and `merge/clean@synth-10k` at 16.5 MB instead of 21.8 MB. On the shipped `dist` build of `grit-cli` alone (one round of 15), the same three rows peak at 16.0 MB instead of 18.5 MB, 17.6 MB instead of 20.2 MB and 15.9 MB instead of 18.0 MB.

```bash
./target/release/grit-bench daily --only pick/one@synth-10k,pick/one@synth-100k,revert/one@synth-10k,merge/clean@synth-10k,merge/clean@synth-100k,merge/content-both-sides@wide-20k,rebase/onto@synth-10k,history/reword@synth-10k \
  --min-runs 15 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-pick-one-6676-after.json
./target/release/grit-bench daily --only status/clean@synth-10k,status/dirty@synth-10k,status/renamed@synth-10k,status/clean@small,diff/worktree@synth-10k,show/head@synth-10k,log/oneline-100@synth-100k,branch/list@many-refs,branch/merged@many-refs,branch/create@many-refs,commit/small@synth-10k,switch/branch@synth-10k,switch/branch@synth-100k,stash/push-pop@synth-10k,add/modified@synth-10k,restore/paths@synth-10k,config/get@small,reset/hard@synth-10k \
  --min-runs 15 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-pick-one-6676-related-after.json
```

## `branch --merged` is start-up and code layout (2026-11-02)

`branch/merged@many-refs` runs `grit branch --merged main` in a clone with 10,000 packed refs: one local branch (`main`), 6,668 remote-tracking branches and 3,333 tags. It prints one line. The daily census had grit at 1.74× Git's time and 2.18× its peak RSS, but the gap is not in the branch query. Under callgrind grit runs 2.27 million instructions and Git 1.09 million, yet `grit --version` alone runs 1.81 million against Git's 0.49 million. The query itself (repository open, ref listing, the reachability check, output) is 0.47 million instructions in grit and 0.59 million in Git, and grit makes fewer system calls (332 against 496). Start-up accounts for the difference: the dynamic loader applies 41,371 relative relocations (0.83 million instructions and 159 copy-on-write faults on `.data.rel.ro`, see [the start-up relocation study](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#startup-relocation-comparison)), clap builds and frees the command tree, and the Rust runtime reads `/proc/self/maps` to find the main thread's stack (0.11 million).

The memory gap is code pages. The heap peaks at 180 KB, and the rest of the resident set is file pages of the binary. The [hot text layout](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#hot-text-layout) had drifted again since it was last regenerated: about 90 commits renamed or added functions on the ref-listing and object-read paths (the pack store's thread-local snapshots, ref store lookups, config overrides) and in CLI start-up (`parse_command`, `normalize_diff_argv`). The command ran code from 68 fault-around windows of 64 KB, 23 of them in plain `.text`. Both fragments were regenerated from the `dist` and `release` builds with the same workload set. The command now touches 41 code windows, 4 of them in plain `.text`. On a `release` build its peak RSS drops from about 10.0 to 8.5 MB, and its time from 1.87 to 1.81 ms (hyperfine medians, three alternating rounds of 1,500 runs).

Rows from `grit-bench daily` on `dist` builds against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0. "base" is `e1e84dfca`. Each cell is the mean of two alternating rounds' medians (30 runs per round), and peak RSS is the mean of the rounds' peaks; the harness runs each command through `sh`. Values are **milliseconds / MB**. Reports (first round): `daily-branch-merged-6636-before.json` → `daily-branch-merged-6636-after.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | time vs git | RSS vs git |
| --- | ---: | ---: | ---: | ---: | ---: |
| `branch/merged@many-refs` | 1.43 / 7.7 | 1.93 / 10.6 | 1.95 / **9.9** | 1.35 → 1.37 | 1.38× → 1.29× |
| `branch/list@many-refs` | 1.30 / 7.0 | 1.73 / 9.5 | 1.78 / **9.2** | 1.33 → 1.37 | 1.36× → 1.31× |
| `branch/list-json@many-refs` | 1.50 / 7.7 | 1.94 / 9.9 | 1.98 / **9.5** | 1.29 → 1.32 | 1.28× → 1.23× |
| `branch/list@many-refs-loose` | 1.33 / 7.0 | 1.76 / 9.6 | 1.83 / **9.3** | 1.32 → 1.37 | 1.37× → 1.33× |
| `branch/create@many-refs` | 1.55 / 7.8 | 2.25 / 10.7 | 2.22 / **10.4** | 1.46 → 1.43 | 1.37× → 1.33× |
| `branch/delete@many-refs` | 4.45 / 8.0 | 3.26 / 11.3 | 3.39 / **10.8** | 0.73 → 0.76 | 1.40× → 1.35× |
| `branch/rename@many-refs` | 1.72 / 7.8 | 3.38 / 10.0 | 3.52 / 10.0 | 1.96 → 2.04 | 1.28× → 1.28× |
| `branch/merged-ref@many-refs` | 0.79 / 6.9 | 1.82 / 8.5 | 1.80 / 8.5 | 2.30 → 2.28 | 1.22× → 1.23× |
| `status/clean@synth-10k` | 6.95 / 9.7 | 6.99 / 13.6 | 7.45 / 13.9 | 1.00 → 1.07 | 1.40× → 1.43× |
| `status/dirty@synth-10k` | 10.56 / 9.8 | 7.27 / 14.2 | 7.60 / **13.8** | 0.69 → 0.72 | 1.44× → 1.41× |
| `diff/worktree@synth-10k` | 6.85 / 10.0 | 6.43 / 13.1 | 6.73 / **12.0** | 0.94 → 0.98 | 1.32× → 1.20× |
| `add/modified@synth-10k` | 16.72 / 12.5 | 10.05 / 18.0 | 9.92 / **17.5** | 0.60 → 0.59 | 1.44× → 1.40× |
| `commit/small@synth-10k` | 27.20 / 14.0 | 16.34 / 17.0 | 16.23 / **16.5** | 0.60 → 0.60 | 1.21× → 1.18× |
| `switch/branch@synth-10k` | 22.50 / 10.7 | 15.89 / 14.0 | 16.55 / **13.4** | 0.71 → 0.74 | 1.31× → 1.25× |
| `merge/clean@synth-10k` | 100.32 / 14.3 | 52.23 / 18.9 | 51.79 / **18.2** | 0.52 → 0.52 | 1.32× → 1.27× |
| `pick/one@synth-10k` | 114.04 / 9.4 | 52.76 / 19.8 | 53.31 / **19.0** | 0.46 → 0.47 | 2.10× → 2.02× |
| `stash/push-pop@synth-10k` | 205.26 / 18.6 | 126.87 / 19.0 | 131.92 / **18.4** | 0.62 → 0.64 | 1.02× → 0.99× |
| `pull/rebase@synth-10k` | 72.63 / 42.3 | 18.43 / 22.8 | 18.70 / **22.5** | 0.25 → 0.26 | 0.54× → 0.53× |
| `clone/local@synth-10k` | 978.14 / 37.3 | 868.81 / 30.7 | 865.02 / 31.0 | 0.89 → 0.88 | 0.82× → 0.83× |

Times moved within the run-to-run noise of this VM. The second round ran slower for both tools: `diff/worktree` read 6.40 → 6.19 ms in round one and 6.45 → 7.26 ms in round two, with Git at 6.57 and 7.40 ms in the same rounds. Measured directly under `perf stat` (150 runs, three alternating rounds), `grit status` on the 10,000-file fixture takes 737 page faults instead of 756 at the same time, and on the many-refs clone 320 instead of 336. `branch/merged-ref` and `branch/rename` time `grit-bench` drivers that link only the library fragment.

Two other start-up changes were measured and left out. Leaking the parsed clap command tree instead of freeing it saves 59,000 instructions, but the allocations that would have reused that memory touch 12 more heap pages, and wall time did not move. A non-PIE link (`-C relocation-model=static`) removes every relative relocation and the 159 copy-on-write faults: on a `release` build `grit --version` drops from 1.4 to 0.93 ms, and `grit branch --merged main` from 1.9 to 1.5 ms and from 9.8 to 8.6 MB. It also gives up ASLR for the executable, which the start-up policy keeps.

```bash
./target/release/grit-bench daily --only branch/merged@many-refs,branch/list@many-refs,branch/list-json@many-refs,branch/list@many-refs-loose,branch/create@many-refs,branch/delete@many-refs,branch/rename@many-refs,branch/merged-ref@many-refs,status/clean@synth-10k,status/dirty@synth-10k,add/modified@synth-10k,commit/small@synth-10k,diff/worktree@synth-10k,switch/branch@synth-10k,merge/clean@synth-10k,pick/one@synth-10k,stash/push-pop@synth-10k,pull/rebase@synth-10k,clone/local@synth-10k \
  --min-runs 30 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-branch-merged-6636-after.json
```

## Commit previews compare object ids only (2026-11-02)

`commit/dry-run@synth-10k` previews a commit of 100 modified files in a 10,000-file clone (`grit commit --dry-run`, against `git commit --dry-run -a`). Grit took twice as long as Git, and longer than the real commit it previewed (`commit/small@synth-10k`, 16 ms). Two leftovers from when the preview diffed trees with rename detection did the extra work:

- **Every index entry was looked up in the object store.** Before listing files, the preview put the object store in its in-memory mode and checked that each of the 10,000 staged blobs existed, re-hashing any it could not find. A blob staged by the preview is never written, so the 100 modified files were read and hashed a second time, and every other path cost a failed loose-object `stat` plus a pack-index lookup: 9,961 failed `statx` calls per run. Nothing after that step reads blob content, because the file list compares object ids and modes. The step is gone.
- **The file list flattened the whole parent tree.** It read all 102 trees of `HEAD` and put every path into a map of owned strings, 10,000 allocations to find 100 changes. It now comes from [`Diff::index_to_tree`](https://docs.rs/grit-lib/latest/grit_lib/diff/struct.Diff.html#method.index_to_tree), which walks the parent tree against the index and skips each directory whose cache-tree entry is still valid, so it reads only the three trees on the changed path (`/`, `data`, `data/d0000`). A root commit lists the index directly. The preview also skips the separate empty-commit check, which ran the same diff a second time; it rejects an empty commit from its own file list with the same error.

That diff pairs and reports paths as lossy UTF-8, while commit file lists use the stored form, which hex-encodes names that are not valid UTF-8. When an index path is not valid UTF-8, or the diff reports a path the lossy form cannot represent, the preview compares raw path bytes against the whole parent tree instead, the old cost. Checking the index paths adds about 1.5% to the preview's instructions on this fixture. Writing a test for that case turned up a staging bug the removed blob check had hidden: dry-run staging handled a modified or deleted tracked file with a non-UTF-8 name under its encoded path, so the preview failed with an I/O error. Staging now decodes the path, and the preview lists those files as the commit does.

Under valgrind the preview runs 51.7 million instructions instead of 165.9 million, and makes 12,611 system calls instead of 22,653. Its heap peaks at 1.7 MB instead of 4.4 MB, with 39,000 allocations instead of 161,000. The rest of its resident set is code and read-only data from the binary: `/usr/bin/time` puts `grit` alone at 10.9 MB instead of 14.0 MB, against Git's 7.9 MB.

Rows from `grit-bench daily` on `dist` builds against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0, page cache dropped before the session. "base" is `d4395a836`. Times are medians of all samples from two alternating rounds (20 runs per round for the target, 15 for the related rows), and peak RSS is the median of the rounds' peaks; the harness runs each command through `sh`, so these peaks are higher than for `grit` alone. Values are **milliseconds / MB**. Reports (first round): `daily-commit-dry-run-6560-before.json` → `daily-commit-dry-run-6560-after.json` and `daily-commit-dry-run-6560-related-before.json` → `daily-commit-dry-run-6560-related-after.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | time vs git | RSS vs git |
| --- | ---: | ---: | ---: | ---: | ---: |
| `commit/dry-run@synth-10k` | 10.73 / 14.6 | 21.32 / 16.6 | **6.84** / **14.1** | 1.99 → 0.64 | 1.14× → 0.96× |
| `add/dry-run@synth-10k` | 7.14 / 13.2 | 9.09 / 14.7 | 8.98 / 15.0 | 1.27 → 1.26 | 1.11× → 1.13× |
| `commit/paths@synth-10k` | 20.88 / 16.8 | 20.64 / 20.3 | 21.14 / 20.3 | 0.99 → 1.01 | 1.21× → 1.21× |
| `commit/small@small` | 4.92 / 12.2 | 4.15 / 11.8 | 4.18 / 11.8 | 0.84 → 0.85 | 0.97× → 0.97× |
| `commit/small@synth-100k` | 895.07 / 38.1 | 123.84 / 33.4 | 123.25 / 33.4 | 0.14 → 0.14 | 0.88× → 0.88× |
| `commit/small@synth-10k` | 26.27 / 14.0 | 15.85 / 17.2 | 15.22 / 16.6 | 0.60 → 0.58 | 1.23× → 1.19× |
| `commit/staged@synth-100k` | 71.36 / 38.3 | 72.20 / 33.2 | 71.48 / 33.1 | 1.01 → 1.00 | 0.87× → 0.86× |
| `status/clean@synth-10k` | 6.51 / 9.7 | 6.29 / 12.8 | 6.30 / 12.2 | 0.97 → 0.97 | 1.32× → 1.26× |
| `status/dirty@synth-10k` | 9.97 / 9.8 | 6.51 / 12.8 | 6.61 / 12.8 | 0.65 → 0.66 | 1.32× → 1.32× |

Apart from the preview, a real commit gained one boolean test, and staging decodes each changed tracked path, which borrows the string for UTF-8 names. The other rows moved within noise. Under valgrind, `grit commit -- <100 paths>` (`commit/paths`) runs 160.6–161.1 million instructions on either build, and `commit/small@small` 7.89 million on either build, within 0.1%. Peak RSS of `commit/small@synth-10k` and `status/*@synth-10k` is the largest of their runs and varies by run for either build, as described below.

```bash
./target/release/grit-bench daily --only commit/dry-run@synth-10k \
  --min-runs 20 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-commit-dry-run-6560-after.json
./target/release/grit-bench daily --only commit/paths@synth-10k,commit/small@small,commit/small@synth-100k,commit/small@synth-10k,commit/staged@synth-100k,add/dry-run@synth-10k,status/clean@synth-10k,status/dirty@synth-10k \
  --min-runs 15 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-commit-dry-run-6560-related-after.json
```

## Small switches join the hot text layout (2026-11-02)

`switch/branch@synth-10k` switches between two branches 50 files apart on a 10,000-file clone. Grit was faster than Git but peaked about 5 MB higher on a `release` build and 3 MB higher on `dist`. Its heap peaks at 1.4 MB, mostly the parsed index, which Git holds as well. The rest was the binary's own code: on `release`, 6.5 MB of the executable segment was resident when `grit switch` exited. The [hot text layout](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#hot-text-layout) had two blind spots for this command:

- **The profiled switch never took the serial write path.** The generator's switch workload changes about 300 files, above `checkout.thresholdForParallelism` (100), so the checkout always used its parallel writer. A switch of fewer files writes them on the caller's thread, and that path was never profiled. On top of that, renames and new code since the last regeneration (pack thread-local snapshots, reference-transaction hooks, checkout write helpers) had dropped out of the lists. Together, about 45 functions with 51 KB of code ran from plain `.text` on a `release` build, each in its own 64 KB fault-around window: 36 windows, about 2.3 MB. The generator now also switches to a branch 50 files away from `main`.
- **Switch sits late in the Gray code order.** Later workloads split into more runs, so switch's 1.3 MB of library code touched 45 of the section's windows. The Gray order is now only the starting point of a seeded search: it moves groups of functions that the same workloads run, and keeps a move when the total windows over all workloads and both builds do not rise and no workload rises above its starting count. On the profile that cut the library section's windows by 6 to 7% and the CLI section's by 10 to 13% across workloads. Switch, refs, branch, show, revs, resolve and stash gained most.

On the 10,000-file clone, the switch's resident code at exit drops from 6.5 to 3.6 MB on `release` and from 4.2 to 2.8 MB on `dist`. In direct runs, the command's peak drops from 13.3 to 11.1 MB on `release` and from 10.8 to 10.1 MB on `dist` (Git: 8.0 MB). Steady-state `grit status` on the same clone drops from 9.6 to 8.8 MB on `dist`. Run time is unchanged: 200 interleaved pairs each put `grit status` at 6.91 → 6.89 ms, `grit config core.bare` at 1.44 → 1.44 ms and the switch pair at 16.03 → 15.95 ms (`release`).

`status/dirty@synth-10k` and `status/renamed@synth-10k` peaks are bimodal, as described in the next section. Round 1 of the change hit the racy path for `status/dirty` (12.9 MB) and round 2 did not (11.0 MB, against 11.7 and 11.8 MB for the base), so the median of the two rounds in the table sits between the modes.

Rows from `grit-bench daily` on `dist` builds against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0. "base" is `765230dc0`. Times are medians of all samples from two alternating rounds of 15 runs, and peak RSS is the median of the rounds' peaks. Values are **milliseconds / MB**. Reports (first round): `daily-switch-branch-10k-6598-before.json` → `daily-switch-branch-10k-6598-after.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | time vs git | RSS vs git |
| --- | ---: | ---: | ---: | ---: | ---: |
| `switch/branch@synth-10k` | 21.51 / 10.2 | 15.79 / 13.1 | **15.09** / **12.4** | 0.73 → 0.70 | 1.29× → 1.22× |
| `switch/branch@synth-100k` | 149.19 / 35.9 | 85.79 / 30.5 | 82.39 / 30.4 | 0.58 → 0.55 | 0.85× → 0.85× |
| `switch/delete-heavy@wide-48k` | 112.85 / 21.7 | 115.14 / 24.5 | 112.30 / 24.0 | 1.02 → 1.00 | 1.13× → 1.10× |
| `switch/large-blob@big-asset` | 864.15 / 307.0 | 741.84 / 13.4 | 745.09 / 12.7 | 0.86 → 0.86 | 0.04× → 0.04× |
| `switch/sparse@sparse-120k` | 40.78 / 7.8 | 20.96 / 12.9 | 21.65 / 12.3 | 0.51 → 0.53 | 1.66× → 1.59× |
| `switch/sparse@sparse-240k` | 60.08 / 8.1 | 20.91 / 12.9 | 21.18 / 12.3 | 0.35 → 0.35 | 1.60× → 1.53× |
| `switch/sparse@sparse-unique-120k` | 74.57 / 17.5 | 28.12 / 15.9 | 28.90 / 15.2 | 0.38 → 0.39 | 0.91× → 0.87× |
| `add/modified@synth-10k` | 15.87 / 11.9 | 9.76 / 17.0 | 9.53 / 16.2 | 0.61 → 0.60 | 1.43× → 1.36× |
| `branch/create@many-refs` | 1.57 / 7.4 | 2.27 / 10.0 | 2.23 / 9.3 | 1.45 → 1.42 | 1.36× → 1.26× |
| `branch/list@many-refs` | 1.32 / 6.6 | 1.82 / 8.9 | 1.68 / 8.5 | 1.37 → 1.27 | 1.35× → 1.30× |
| `commit/dry-run@synth-10k` | 10.89 / 13.9 | 31.27 / 16.0 | 31.14 / 15.4 | 2.87 → 2.86 | 1.15× → 1.11× |
| `commit/small@synth-10k` | 25.70 / 13.1 | 15.70 / 15.9 | 15.78 / 14.2 | 0.61 → 0.61 | 1.21× → 1.09× |
| `config/get@small` | 1.21 / 6.4 | 1.57 / 7.8 | 1.54 / 7.2 | 1.29 → 1.27 | 1.21× → 1.12× |
| `diff/worktree@synth-10k` | 6.48 / 9.5 | 6.18 / 12.3 | 6.27 / 11.7 | 0.95 → 0.97 | 1.30× → 1.24× |
| `init/empty@small` | 2.41 / 6.6 | 2.21 / 9.5 | 2.27 / 9.3 | 0.92 → 0.94 | 1.44× → 1.40× |
| `log/oneline-100@synth-100k` | 2.23 / 9.2 | 4.22 / 12.8 | 4.11 / 12.0 | 1.90 → 1.85 | 1.40× → 1.30× |
| `merge/clean@synth-10k` | 97.36 / 13.6 | 49.06 / 18.2 | 49.62 / 17.4 | 0.50 → 0.51 | 1.34× → 1.28× |
| `pick/one@synth-10k` | 110.57 / 9.0 | 51.18 / 18.4 | 50.88 / 17.8 | 0.46 → 0.46 | 2.05× → 1.98× |
| `reflog/list@synth-10k` | 3.60 / 9.0 | 31.91 / 9.9 | 32.14 / 9.2 | 8.85 → 8.92 | 1.10× → 1.03× |
| `reset/hard@synth-10k` | 15.01 / 10.6 | 14.10 / 15.3 | 14.01 / 15.1 | 0.94 → 0.93 | 1.44× → 1.43× |
| `restore/paths@synth-10k` | 11.57 / 12.1 | 12.95 / 15.9 | 13.07 / 15.3 | 1.12 → 1.13 | 1.32× → 1.27× |
| `show/head@synth-10k` | 2.00 / 8.6 | 2.50 / 9.9 | 2.41 / 9.6 | 1.25 → 1.20 | 1.16× → 1.12× |
| `stash/push-pop@synth-10k` | 194.12 / 17.6 | 120.98 / 17.8 | 122.78 / 16.9 | 0.62 → 0.63 | 1.01× → 0.96× |
| `status/clean@small` | 1.87 / 7.2 | 2.66 / 9.3 | 2.73 / 8.9 | 1.42 → 1.46 | 1.29× → 1.24× |
| `status/clean@synth-10k` | 6.74 / 9.2 | 6.77 / 13.5 | 6.91 / 13.0 | 1.00 → 1.03 | 1.46× → 1.40× |
| `status/dirty@synth-100k` | 92.36 / 28.9 | 51.14 / 30.9 | 49.83 / 30.7 | 0.55 → 0.54 | 1.07× → 1.06× |
| `status/dirty@synth-10k` | 10.21 / 9.3 | 7.19 / 11.7 | 7.33 / 11.9 | 0.70 → 0.72 | 1.27× → 1.28× |
| `status/renamed@synth-10k` | 6.84 / 9.2 | 6.98 / 13.7 | 6.81 / 13.0 | 1.02 → 1.00 | 1.49× → 1.41× |
| `worktree/list@synth-10k` | 1.52 / 6.4 | 2.38 / 8.6 | 2.43 / 8.3 | 1.57 → 1.60 | 1.34× → 1.30× |

The `release` build shows the same pattern with larger steps, since its lists had drifted further: `switch/branch@synth-10k` 15.6 → 13.3 MB, `status/clean@synth-10k` 15.4 → 12.3 MB, `show/head@synth-10k` 11.9 → 9.6 MB and `stash/push-pop@synth-10k` 20.2 → 17.7 MB, with run times within noise.

```bash
./target/release/grit-bench daily --only switch/branch@synth-10k,switch/branch@synth-100k,switch/delete-heavy@wide-48k,switch/large-blob@big-asset,switch/sparse@sparse-120k,switch/sparse@sparse-240k,switch/sparse@sparse-unique-120k \
  --min-runs 15 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-switch-branch-10k-6598-after.json
```

## Status code layout by workload set (2026-11-01)

`status/clean@synth-10k` runs `grit status` on a clean 10,000-file clone. Grit took as long as Git but peaked 2 to 6 MB higher. Its heap peaks at only 1.7 MB; most of the rest is the binary's own pages, and on a `release` build `grit status` left 6.9 MB of code and read-only data resident. Three things kept the [hot text layout](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#hot-text-layout) from doing its job, and a fourth cost only the first status after a checkout:

- **The lists had drifted.** Commits since the morning's profile renamed or added functions on the status path (the Odb-owned loose negative cache, config override handling, the stat scan's thread count). Their patterns no longer matched, so on a `release` build status ran about 70 functions from plain `.text`, each in its own 64 KB fault-around window: 29 windows, about 1.9 MB.
- **Merged functions kept the wrong name.** LLVM merges functions with identical code (`RawVec<T>::grow_one` for every same-sized `T`) into one body and keeps the other names as aliases. The body's input section is named after one of them, while callgrind reports whichever name it picks for the address, so 15 such bodies stayed in `.text` even with fresh lists. The generator now lists every symbol at an executed function's address.
- **Each command's code was sliced by sharing count.** The lists were sorted by how many workloads run a function, which cut one command's code into a slice per count spread over the whole section: status touched 22 of the library section's 55 windows to run 5% of its code. The library list now groups functions by the exact set of workloads that run them and orders the groups in reflected binary (Gray code) order over the workloads, so neighbouring groups differ in as few workloads as possible. On the profile, the four status workloads touch 16 to 19 library windows instead of 22 to 26, and all 20 workloads together 469 instead of 546; commit and stash, late in the list, touch one to five more. The CLI list keeps the sort by sharing count with the Gray order as tie-break (281 → 258 windows, only stash one more): the full Gray order there cost `grit config` and `grit tag` a window each, 2 to 4% of their run time.
- **The first status after a checkout copied attributes for every file.** When every entry is racy, the stat refresh hashes them all and resolved each entry's attributes into its job first, a few hundred bytes per file. That list peaked at 3.9 MB for 10,000 files. Jobs now point into the distinct attribute sets seen, and that run's heap peak drops from 7.7 to 4.1 MB; on a fresh clone the first `grit status` peaks at 12.9 MB instead of 15.0 MB (Git: 9.4 MB).

On a `release` build `grit status` on the clone now keeps 4.8 MB of its binary resident instead of 6.9 MB, with no function it runs left in plain `.text`; the stripped `dist` build drops from 4.7 to 4.0 MB.

A daily row reports the largest peak of its runs, and the first runs after the harness recreates a fixture take the racy path above, so `status/clean@synth-10k` and `status/renamed@synth-10k` peaks are bimodal for either build. Over the three rounds below, the base peaked at 11.4–11.5 MB or 15.0–15.5 MB and the change at 10.8–11.3 MB or 13.2–13.3 MB; the medians in the table depend on which mode each round hit.

Rows from `grit-bench daily` on `dist` builds against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0, page cache dropped before the session. "base" is `0fef76aed`. Times are medians of all samples from three alternating rounds of 20 runs (the related rows: two rounds of 15), and peak RSS is the median of the rounds' peaks. Values are **milliseconds / MB**. Reports (first round): `daily-status-clean-10k-6351-before.json` → `daily-status-clean-10k-6351-after.json` and `daily-status-clean-10k-6351-related-before.json` → `daily-status-clean-10k-6351-related-after.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | time vs git | RSS vs git |
| --- | ---: | ---: | ---: | ---: | ---: |
| `status/clean@synth-10k` | 6.91 / 9.6 | 6.65 / 11.5 | **6.64** / **11.3** | 0.96 → 0.96 | 1.20× → 1.18× |
| `status/clean@deep-tree` | 1069.54 / 40.5 | 499.02 / 39.3 | 504.52 / 39.2 | 0.47 → 0.47 | 0.97× → 0.97× |
| `status/clean@small` | 1.93 / 7.4 | 2.60 / 9.7 | 2.56 / 9.2 | 1.35 → 1.33 | 1.31× → 1.23× |
| `status/clean@synth-100k` | 60.17 / 30.3 | 48.13 / 31.8 | 45.83 / 31.7 | 0.80 → 0.76 | 1.05× → 1.05× |
| `status/dirty@deep-tree` | 1128.15 / 41.3 | 537.27 / 46.7 | 543.83 / 46.2 | 0.48 → 0.48 | 1.13× → 1.12× |
| `status/dirty@small` | 1.86 / 7.4 | 2.68 / 9.8 | 2.51 / 9.2 | 1.44 → 1.35 | 1.31× → 1.24× |
| `status/dirty@synth-100k` | 95.48 / 30.2 | 52.86 / 31.8 | 48.37 / 31.7 | 0.55 → 0.51 | 1.05× → 1.05× |
| `status/dirty@synth-10k` | 10.39 / 9.7 | 7.01 / 11.6 | 6.93 / 10.9 | 0.68 → 0.67 | 1.20× → 1.13× |
| `status/paths@synth-100k` | 44.70 / 30.2 | 42.77 / 32.1 | 41.73 / 31.9 | 0.96 → 0.93 | 1.06× → 1.06× |
| `status/renamed@synth-10k` | 6.88 / 9.6 | 6.80 / 15.0 | 6.38 / 10.8 | 0.99 → 0.93 | 1.55× → 1.12× |
| `add/modified@synth-10k` | 16.39 / 12.4 | 9.77 / 17.5 | 9.69 / 17.0 | 0.60 → 0.59 | 1.41× → 1.37× |
| `blame/file@git-git` | 272.33 / 243.8 | 228.11 / 264.8 | 227.22 / 264.8 | 0.84 → 0.83 | 1.09× → 1.09× |
| `branch/create@many-refs` | 1.58 / 7.9 | 2.17 / 9.9 | 2.18 / 9.7 | 1.37 → 1.37 | 1.25× → 1.23× |
| `branch/list@many-refs` | 1.33 / 6.8 | 1.75 / 9.0 | 1.77 / 8.8 | 1.32 → 1.33 | 1.33× → 1.29× |
| `commit/small@synth-10k` | 26.48 / 13.8 | 15.47 / 16.2 | 15.29 / 17.0 | 0.58 → 0.58 | 1.18× → 1.23× |
| `config/get@small` | 1.17 / 6.8 | 1.56 / 7.7 | 1.58 / 7.9 | 1.34 → 1.35 | 1.13× → 1.16× |
| `diff/worktree@synth-10k` | 6.69 / 9.9 | 6.12 / 12.4 | 5.96 / 12.4 | 0.91 → 0.89 | 1.26× → 1.26× |
| `log/json-100@git-git` | 2.87 / 14.9 | 3.77 / 17.3 | 3.71 / 16.7 | 1.32 → 1.29 | 1.16× → 1.12× |
| `log/oneline-100@git-git` | 2.81 / 14.9 | 3.22 / 16.8 | 3.21 / 16.3 | 1.14 → 1.14 | 1.13× → 1.09× |
| `merge/clean@synth-10k` | 98.25 / 14.2 | 49.92 / 18.3 | 50.31 / 17.8 | 0.51 → 0.51 | 1.29× → 1.26× |
| `notes/list@synth-10k` | 1.27 / 6.8 | 1.90 / 9.4 | 1.93 / 9.4 | 1.50 → 1.52 | 1.39× → 1.39× |
| `reset/mixed@synth-10k` | 7.93 / 9.7 | 9.64 / 16.2 | 9.32 / 12.6 | 1.22 → 1.18 | 1.67× → 1.31× |
| `resolve/head-n@git-git` | 1.18 / 6.8 | 1.81 / 8.9 | 1.80 / 8.7 | 1.53 → 1.52 | 1.31× → 1.28× |
| `revs/count@git-git` | 493.27 / 62.8 | 618.11 / 124.1 | 620.05 / 123.9 | 1.25 → 1.26 | 1.98× → 1.97× |
| `show/head@synth-10k` | 2.01 / 8.9 | 2.36 / 10.1 | 2.40 / 10.2 | 1.17 → 1.19 | 1.14× → 1.16× |
| `stash/push-pop@synth-10k` | 192.48 / 18.4 | 115.99 / 18.4 | 115.28 / 17.8 | 0.60 → 0.60 | 1.00× → 0.97× |
| `switch/branch@synth-10k` | 21.47 / 10.7 | 15.27 / 13.5 | 15.30 / 13.0 | 0.71 → 0.71 | 1.26× → 1.22× |
| `tag/list@many-refs` | 3.21 / 8.1 | 3.16 / 9.9 | 3.23 / 9.7 | 0.98 → 1.01 | 1.22× → 1.20× |
| `worktree/list@synth-10k` | 1.55 / 6.8 | 2.31 / 8.3 | 2.31 / 8.4 | 1.48 → 1.49 | 1.22× → 1.23× |

Run time moved within noise. Interleaved direct runs (prepare-free, 200 to 500 pairs each) put `grit status` on the synth-10k clone at 5.75 → 5.67 ms, `grit tag` on many-refs at 2.59 → 2.57 ms, `grit branch` at 1.16 → 1.14 ms and `grit config core.repositoryformatversion` at 0.989 → 0.992 ms. `config/get@small` keeps one more library window resident (about 60 KB of its 3.2 MB of binary pages); its row's peak is the largest of runs that spread from 6.4 to 7.6 MB for either build. `commit/small` peaks are bimodal the same way as status, from 14.5 to 17.7 MB for either build across sessions.

```bash
./target/release/grit-bench daily --only status/clean@synth-10k,status/clean@deep-tree,status/clean@small,status/clean@synth-100k,status/dirty@deep-tree,status/dirty@small,status/dirty@synth-100k,status/dirty@synth-10k,status/paths@synth-100k,status/renamed@synth-10k \
  --min-runs 20 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-status-clean-10k-6351-after.json
```

## Packing refs merges into packed-refs in one pass (2026-11-01)

`gc/pack-refs@synth-10k` packs one new loose tag into a `packed-refs` file of 935 records (`grit gc pack-refs`, against `git pack-refs`). On the `dist` build grit took 3.30 ms and 10.5 MB against Git's 1.72 ms and 7.1 MB. The files backend validated `packed-refs`, discarded the result, then parsed the file a second time into a `HashMap<String, _>`, copied and sorted every name as an owned `String`, and re-encoded every object id to write the file back. To decide whether each loose ref needed a peel line it inflated the whole object it named, through a 256 KiB scratch buffer that was mapped and unmapped on every run, and the two lock timeouts loaded the configuration from disk twice more.

Packing now collects the loose candidates first, using the directory entries' types instead of two `stat` calls per entry and a header-only kind lookup instead of a full read. It then parses `packed-refs` once, into records that borrow their names from the file bytes, and merges the sorted loose refs in by name. A file that is not sorted, or repeats a name, is still sorted with the last record winning. The `packed-refs` lock is now taken before the file is read, as Git does, so a concurrent rewrite cannot be lost between the read and the rename, and the lock timeouts come from the repository's loaded configuration. Measured directly with 200 alternating spawns on a `release` build, `gc pack-refs` went from 2.7 ms to 2.3 ms (Git 1.5 ms), with 30% fewer page faults and system time down from 1.1 ms to 0.7 ms.

What remains of the gap is mostly process start-up, which every command pays: `grit --version` alone costs about 1.2 ms and 5.4 MB on `dist`, most of it the copy-on-write faults on `.data.rel.ro` described in [the start-up relocation study](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#startup-relocation-comparison). Pack-refs code is also not part of the [hot text layout](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#hot-text-layout), whose workloads do not include `gc`, so it faults in about 13 windows of plain `.text` that `--version` does not touch.

Rows from `grit-bench daily` on `dist` builds against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0. "base" is `3344230ec`. Times are medians of all samples from two alternating rounds (20 runs per round; the clone, `gc/default` and `remote/*` rows get fewer within the time budget), and peak RSS is the median of the rounds' peaks. Values are **milliseconds / MB**. Reports (first round): `daily-pack-refs-6454-before.json` → `daily-pack-refs-6454-after.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | time vs git | RSS vs git |
| --- | ---: | ---: | ---: | ---: | ---: |
| `gc/pack-refs@synth-10k` | 1.72 / 7.1 | 3.30 / 10.5 | 2.85 / 10.4 | 1.91 → 1.65 | 1.48× → 1.46× |
| `gc/default@synth-10k` | 345.11 / 49.7 | 276.94 / 50.4 | 270.14 / 50.8 | 0.80 → 0.78 | 1.01× → 1.02× |
| `clone/bare@synth-10k` | 858.13 / 37.2 | 715.64 / 24.7 | 710.87 / 24.4 | 0.83 → 0.83 | 0.66× → 0.66× |
| `clone/local@synth-10k` | 999.18 / 37.3 | 854.14 / 28.6 | 852.32 / 29.8 | 0.85 → 0.85 | 0.77× → 0.80× |
| `clone/shallow-depth1@synth-10k` | 202.99 / 32.3 | 205.10 / 26.4 | 187.88 / 26.3 | 1.01 → 0.93 | 0.82× → 0.81× |
| `branch/delete@many-refs` | 4.59 / 8.1 | 3.93 / 10.6 | 3.59 / 10.7 | 0.86 → 0.78 | 1.32× → 1.32× |
| `tag/create@many-refs` | 1.56 / 7.7 | 2.44 / 10.6 | 2.59 / 10.9 | 1.56 → 1.66 | 1.37× → 1.41× |
| `remote/remove@many-refs` | 831.93 / 10.1 | 217.77 / 23.7 | 219.39 / 23.8 | 0.26 → 0.26 | 2.34× → 2.35× |
| `remote/rename@many-refs` | 56677.16 / 9.4 | 3549.41 / 83.4 | 3690.33 / 83.5 | 0.06 → 0.07 | 8.86× → 8.86× |

Clones pack their imported refs through the same code, and ref deletions share the rewritten lock helper. The swings in the other rows are noise, checked directly. `grit tag` on the many-refs fixture measures 1.57 ms of task-clock with 333–334 page faults on both builds over 300 alternating runs. `remote rename` measures 4.05 s against 4.09 s when the after build runs first (run before-first, it reads 10% slower, because the fixture slows with every rename cycle). An HTTP clone of synth-10k has the same 6.25 MB heap peak on both builds, with 2,300 fewer allocations after, so the `clone/local` RSS step is in file-backed pages.

```bash
./target/dist/grit-bench daily --only gc/pack-refs@synth-10k,gc/default@synth-10k,clone/bare@synth-10k,clone/local@synth-10k,clone/shallow-depth1@synth-10k,branch/delete@many-refs,remote/remove@many-refs,remote/rename@many-refs,tag/create@many-refs \
  --min-runs 20 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-pack-refs-6454-after.json
```

## One-line log reads commits only through their subject (2026-11-01)

`log/oneline-100@git-git` lists the last 100 commits of git.git one line each (`grit log --limit 100`), against `git log --oneline -100`. On the `dist` build grit took 3.50 ms against Git's 2.94 ms. The walk reads about 140 commits from the pack to order them by date, and inflated every one in full, keeping the bytes for the output rows. Those commits hold 115 KB of text, but a one-line row needs only the headers (parents and dates for the walk; author, date and change id for the row) and the subject: 40 KB. The other two thirds were message bodies inflated and then dropped.

[`Odb::read_commit_through_subject`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/struct.Odb.html#method.read_commit_through_subject) returns just that prefix. For a commit the pack stores whole, it feeds the inflater 128 compressed bytes at a time and stops as soon as the first non-blank message line is complete. Stepping through the input rather than the output keeps zlib in its fast decoding loop, which needs 258 bytes of output room; an earlier cut that stepped the output 512 bytes at a time decoded half of every step on the slow path and was slower than inflating everything. The commit stream takes the extent to keep ([`CommitExtent::ThroughSubject`](https://docs.rs/grit-lib/latest/grit_lib/revwalk/enum.CommitExtent.html)), and human and Markdown `log` and `revs` listings ask for it. JSON rows, decorations and diff formats still read whole commits. The new functions joined the [hot text layout](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#hot-text-layout), which was regenerated; without that they faulted in about 390 KiB more code on this listing.

Per commit, grit's walk now costs less than Git's: from `--limit 1` to `--limit 100` grit adds 1.35 ms and Git 1.51 ms. What remains of the gap is process start-up: `grit --version` takes 1.24 ms against Git's 0.88 ms, mostly loader relocations and the copy-on-write faults on `.data.rel.ro` described in [the start-up relocation study](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#startup-relocation-comparison). Measured directly with 600 alternating spawns, the median goes from 3.29 ms to 2.94 ms against Git's 2.70–2.85 ms, and `/usr/bin/time` peak RSS from 14.45 MB to 14.26 MB (Git 12.6 MB).

Rows from `grit-bench daily` on `dist` builds against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0. "base" is `3edaa6f42`. Times are medians of all samples from two alternating rounds (20 runs per round), and peak RSS is the median of the rounds' peaks. Values are **milliseconds / MB**. Reports (first round): `daily-log-oneline-100-6310-before.json` → `daily-log-oneline-100-6310-after.json`; the headline census with the base build as `baseline_grit` is `daily-log-oneline-100-6310-headlines.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | time vs git | RSS vs git |
| --- | ---: | ---: | ---: | ---: | ---: |
| `log/all@git-git` | 45.14 / 20.9 | 46.58 / 27.5 | 39.41 / 26.7 | 1.03 → 0.87 | 1.31× → 1.28× |
| `log/author-n50@deep-history` | 1.73 / 9.1 | 2.87 / 13.6 | 2.96 / 13.6 | 1.66 → 1.71 | 1.50× → 1.50× |
| `log/author-n50@deep-history-cg` | 2.00 / 11.7 | 2.96 / 16.2 | 3.01 / 16.3 | 1.48 → 1.51 | 1.38× → 1.39× |
| `log/author@git-git` | 2.41 / 12.0 | 3.79 / 16.8 | 4.06 / 17.1 | 1.57 → 1.68 | 1.40× → 1.42× |
| `log/json-100@git-git` | 2.91 / 15.1 | 3.82 / 17.4 | 3.71 / 17.2 | 1.31 → 1.28 | 1.15× → 1.14× |
| `log/json-filter@git-git` | 2.28 / 12.2 | 3.78 / 17.0 | 4.01 / 17.2 | 1.66 → 1.75 | 1.39× → 1.41× |
| `log/json-n10-loose-tags@git-git` | 12.31 / 17.4 | 12.08 / 20.0 | 12.24 / 19.9 | 0.98 → 0.99 | 1.15× → 1.14× |
| `log/json-n10-packed-tags@git-git` | 10.06 / 17.4 | 2.56 / 11.2 | 2.50 / 10.8 | 0.25 → 0.25 | 0.64× → 0.62× |
| `log/limit-path@git-git` | 19.04 / 33.6 | 14.19 / 33.6 | 15.46 / 33.7 | 0.75 → 0.81 | 1.00× → 1.00× |
| `log/oneline-100@git-git` | 2.94 / 14.9 | 3.50 / 16.8 | 3.20 / 16.8 | 1.19 → 1.09 | 1.13× → 1.13× |
| `log/oneline-100@synth-100k` | 2.19 / 10.0 | 2.54 / 11.5 | 2.48 / 11.9 | 1.16 → 1.13 | 1.16× → 1.19× |
| `log/path@git-git` | 349.61 / 261.9 | 377.69 / 277.6 | 353.77 / 277.7 | 1.08 → 1.01 | 1.06× → 1.06× |
| `log/stat-n50@git-git` | 99.34 / 40.6 | 161.24 / 77.0 | 168.51 / 77.0 | 1.62 → 1.70 | 1.89× → 1.90× |
| `revs/author-n50@deep-history` | 1.62 / 9.0 | 2.49 / 11.6 | 2.62 / 12.2 | 1.53 → 1.61 | 1.29× → 1.36× |
| `show/head@git-git` | 1.88 / 9.0 | 2.27 / 10.5 | 2.31 / 10.3 | 1.21 → 1.23 | 1.16× → 1.14× |

Only `log/oneline-100` takes the new path; `log/oneline-100@synth-100k` has short messages, so there is little body to skip. The other rows run code this change does not touch apart from one branch on the retained extent. Their swings are run-to-run noise: re-run with 200 to 300 alternating spawns each, `log --author=. -n 50 --json`, `log --json --filter` and `log --limit 100 --json` stay within 2% of the base build with identical page-fault counts, `perf stat` puts `log --author` on either side of the base build from round to round, and `log -n 200 -- Documentation` and `log --stat` land within 0.5%. The headline census (status, diff, add, commit, switch, branch, merge, pick, stash, fetch, pull, push, clone, show, blame) moved within noise; peak RSS of `status`, `diff`, `show` and `branch` is flat or lower with the regenerated layout.

```bash
./target/release/grit-bench daily --only log/all@git-git,log/author-n50@deep-history,log/author-n50@deep-history-cg,log/author@git-git,log/json-100@git-git,log/json-filter@git-git,log/json-n10-loose-tags@git-git,log/json-n10-packed-tags@git-git,log/limit-path@git-git,log/oneline-100@git-git,log/oneline-100@synth-100k,log/path@git-git,log/stat-n50@git-git,revs/author-n50@deep-history,show/head@git-git \
  --min-runs 20 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-log-oneline-100-6310-after.json
```

## Streamed JSON log listing (2026-11-01)

`log/json-100@git-git` lists the last 100 commits of git.git as JSON (`grit log --limit 100 --json`), against `git log -100` with a verbose `--format`. Git ran in 2.97 ms and 15.0 MB on the `dist` build, grit in 4.43 ms and 18.7 MB. Inflating the 131 commits the walk reads costs both tools about the same. The rest of the gap was work grit repeated and code it ran from scattered pages:

- **Decorations parsed `packed-refs` twice.** The JSON rows carry ref decorations, so grit scans the 1,909 packed refs for the ones naming a listed commit. The filtered scan validated the whole file once to prove it well-formed and sorted, then parsed it again to emit records, about 12% of the run. It now makes one strict pass, holds the few records the filter keeps (borrowed from the mapped file) and emits them once the pass succeeds. A corrupt or out-of-order file still emits nothing.
- **The streamed listing was set up twice.** Deciding whether a non-terminal listing can stream discovered the repository, loaded config and built the walk options, and then the listing did it all again (about 7%). The check now hands its repository and walk to the listing, and runs only after the cheap page-size check.
- **Rows were written in fragments.** serde_json wrote each row as dozens of small pieces to the line-buffered stdout, which scanned every piece for a newline (about 5%). Each row is now serialized into memory and written once.
- **The hot code path missed release builds and streamed listings.** The [hot text layout](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#hot-text-layout) was profiled on the fat-LTO `dist` build only. ThinLTO in the `release` build renames functions shared between codegen units to `name.llvm.<n>`, which the bare patterns did not match, and keeps functions apart that `dist` inlines (clap's command construction, repository open, config loading). On a `release` build this listing faulted in 122 windows of plain `.text`, about 7.6 MB resident. Every pattern now also matches the suffixed name, the generator profiles the `dist` and `release` builds and merges them, and a streamed `--json` listing joins the log workloads. Code faults in plain `.text` drop from 122 to 17 on the `release` build.

The 178 copy-on-write faults on `.data.rel.ro` and the pack index pages (which Git touches too) remain.

Rows from `grit-bench daily` on `dist` builds against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0. "base" is `d2551aef4`. Times are medians of all samples from two alternating rounds (20 runs per round for the log rows, 15 for the related rows), and peak RSS is the median of the rounds' peaks. Values are **milliseconds / MB**. Reports (first round): `daily-log-json-100-6206-before.json` → `daily-log-json-100-6206-after.json` and `daily-log-json-100-6206-related-before.json` → `daily-log-json-100-6206-related-after.json`.

| Scenario | git ms / MB | base ms / MB | after ms / MB | time vs git | RSS vs git |
| --- | ---: | ---: | ---: | ---: | ---: |
| `log/json-100@git-git` | 2.97 / 15.0 | 4.43 / 18.7 | **3.79** / **17.0** | 1.49 → 1.27 | 1.24× → 1.13× |
| `log/oneline-100@git-git` | 2.96 / 15.0 | 3.78 / 18.3 | 3.40 / 16.6 | 1.28 → 1.15 | 1.22× → 1.11× |
| `log/oneline-100@synth-100k` | 2.19 / 9.7 | 2.99 / 13.6 | 2.58 / 11.0 | 1.37 → 1.18 | 1.40× → 1.13× |
| `log/json-n10-packed-tags@git-git` | 9.97 / 17.5 | 3.00 / 12.8 | 2.60 / 12.0 | 0.30 → 0.26 | 0.73× → 0.69× |
| `log/json-n10-loose-tags@git-git` | 12.13 / 17.5 | 12.58 / 20.5 | 11.94 / 19.7 | 1.04 → 0.98 | 1.17× → 1.13× |
| `log/json-filter@git-git` | 2.40 / 12.0 | 4.29 / 18.0 | 4.24 / 17.4 | 1.79 → 1.77 | 1.50× → 1.45× |
| `log/author@git-git` | 2.43 / 11.8 | 3.96 / 17.4 | 3.90 / 16.7 | 1.63 → 1.60 | 1.47× → 1.41× |
| `log/author-n50@deep-history` | 1.75 / 9.1 | 2.95 / 14.0 | 2.92 / 13.4 | 1.69 → 1.67 | 1.54× → 1.47× |
| `log/author-n50@deep-history-cg` | 1.98 / 12.2 | 3.08 / 16.5 | 3.05 / 16.1 | 1.55 → 1.54 | 1.36× → 1.33× |
| `log/all@git-git` | 44.72 / 21.0 | 48.92 / 28.8 | 46.15 / 27.3 | 1.09 → 1.03 | 1.37× → 1.30× |
| `log/stat-n50@git-git` | 98.36 / 40.7 | 163.59 / 84.1 | 158.22 / 82.8 | 1.66 → 1.61 | 2.07× → 2.03× |
| `log/path@git-git` | 337.61 / 261.0 | 353.58 / 277.6 | 347.59 / 276.8 | 1.05 → 1.03 | 1.06× → 1.06× |
| `log/limit-path@git-git` | 20.43 / 33.9 | 14.85 / 34.3 | 16.58 / 33.7 | 0.73 → 0.81 | 1.01× → 1.00× |
| `show/head@synth-10k` | 2.04 / 8.9 | 2.80 / 11.7 | 2.68 / 11.4 | 1.37 → 1.31 | 1.31× → 1.27× |
| `revs/author-n50@deep-history` | 1.70 / 8.9 | 2.73 / 13.6 | 2.65 / 13.0 | 1.60 → 1.56 | 1.53× → 1.45× |
| `branch/list-json@many-refs` | 1.55 / 7.7 | 2.04 / 9.4 | 1.88 / 9.6 | 1.31 → 1.21 | 1.22× → 1.24× |
| `tag/list@many-refs` | 3.32 / 8.2 | 3.49 / 10.2 | 3.23 / 10.1 | 1.05 → 0.97 | 1.24× → 1.24× |
| `status/clean@small` | 1.91 / 7.6 | 2.57 / 9.6 | 2.55 / 9.8 | 1.34 → 1.33 | 1.27× → 1.29× |

`log/limit-path` lists through the path-limited walk, where both builds do the same work; `perf stat` over 3 × 30 alternating runs puts it at 14.0 → 14.1 ms of CPU with 6 fewer page faults, so the row's swing is run-to-run noise. The other related rows (status, diff, add, commit, switch, config, blame, `revs/count`) moved within noise; `status/dirty@synth-10k`'s peak RSS is the largest of its runs and swings between 11.4 and 15.3 MB for either build depending on whether a run takes the threaded index refresh.

On `release` builds `log/json-100@git-git` goes from 4.64 ms / 24.8 MB to 4.17 ms / 19.0 MB (git 2.96 ms / 15.0 MB): the layout fix matters most there, because release builds matched far fewer patterns.

```bash
./target/release/grit-bench daily --only log/all@git-git,log/author-n50@deep-history,log/author-n50@deep-history-cg,log/author@git-git,log/json-100@git-git,log/json-filter@git-git,log/json-n10-loose-tags@git-git,log/json-n10-packed-tags@git-git,log/limit-path@git-git,log/oneline-100@git-git,log/oneline-100@synth-100k,log/path@git-git,log/stat-n50@git-git \
  --min-runs 20 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-log-json-100-6206-after.json
```

## Small commit diffs stay on one thread (2026-10-31)

`show`, `diff A B` and `log --stat` load and diff each changed file of a commit as one unit of work. With `core.preloadIndex` on (Git's default), any commit touching two or more files sent those units to the worker pool. On `show/head@synth-10k`, a commit of eight two-line files, that meant three new threads, each with its own stack, allocator arena and pack inflater, to do less work than starting them cost: about 0.3 ms more wall time and 30% more CPU than running serially.

Timing `grit show` on git.git commits with the pool on and off (`core.preloadIndex=false`) put the crossover at about 18 files. Up to 17 files the serial run was as fast or faster (by 0.1–0.35 ms); from 18 files the pool won, taking less than half the serial time at 225 files. Commits with few but very large files gain nothing from the pool either, since each file is diffed by one thread. `grit-lib` now uses the pool for 16 or more entries (or from the caller's `ThreadPolicy::min_items_override`). Starting it two files early costs a few tenths of a millisecond on 16- and 17-file commits, while starting it late would give up larger wins.

Daily rows, mean of two alternating rounds of 30 runs each (`dist` builds, page cache dropped before the session), on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0. "Base" is `d2551aef4`, which already links the CLI hot-text layout (next section). Reports: `daily-show-diff-pool-5993-before.json` → `daily-show-diff-pool-5993-after.json` (round one).

| Scenario | git ms / MiB | base | + change |
| --- | ---: | ---: | ---: |
| `show/head@synth-10k` | 2.00 / 8.5 | 2.66 / 10.8 | **2.38** / **9.5** |
| `show/head@git-git` | 1.94 / 8.4 | 2.21 / 9.6 | 2.26 / 9.8 |
| `show/head@big-blobs` | 1.59 / 7.1 | 2.00 / 9.3 | 1.92 / 9.3 |
| `log/stat-n50@git-git` | 96.36 / 38.8 | 154.48 / 77.0 | 158.29 / **74.6** |
| `diff/worktree@synth-10k` | 7.19 / 9.4 | 7.23 / 11.8 | 6.93 / 12.4 |

`git-git` and `big-blobs` show one-file commits, which were already serial; their rows are noise. `log --stat -n 50` on git.git is mostly merges and one-file commits (three of the fifty touch two files); it peaks about 2.4 MiB lower, and its time is within noise: four interleaved hyperfine series of 40 runs read 158.3 and 151.2 ms for the base against 154.7 and 158.5 ms with the change. `patch-id/commit@git-git` measured 1.8–1.9 ms for both binaries in 300 interleaved hyperfine runs. Commit sizes from git.git (`show` medians of 40 runs, ms, measured on the earlier base `b19e946fe`):

| Files changed | base | + change | git |
| ---: | ---: | ---: | ---: |
| 8 | 4.71 | **4.52** | 4.37 |
| 12 | 6.95 | **6.55** | 7.46 |
| 16 | 8.66 | 8.46 | 9.13 |
| 28 | 8.55 | 8.58 | 14.85 |
| 109 | 8.80 | 8.61 | 18.81 |
| 225 | 19.43 | 19.19 | 37.28 |

Peak RSS comparisons between two builds need the page cache in the same state. A binary that was just copied into place showed about 0.4 MiB more resident code and read-only data than the same file after `echo 3 > /proc/sys/vm/drop_caches`, on identical runs. Both builds above were measured after dropping the cache.

## Restoring hunks of a large file (2026-11-01)

`restore/hunks@large-file` discards 250 of 500 one-line hunks in a 50,000-line, 490 KB `large.txt` inside `synth-10k`. Git only applies a prebuilt patch (`git apply -R`), while grit has to diff the file to resolve the hunk ids, so some gap is inherent. The rest was memory and repeated work. A run faulted in about 3,200 more anonymous pages than a trivial command: the heap peaked at 13.2 MB, almost all of it short-lived line tables of the Myers diff. On top of that, the 490 KB file was hashed with SHA-1 three times.

What changed:

- The line diff keeps one 16-byte record per line (body plus newline) instead of a 24-byte body and ending pair, and the raw compare keys borrow those records instead of filling a second table. Compaction reads the same keys for its indentation measure. The interner is freed before the search, and the K-vectors are a fresh zeroed allocation, so diagonals the search never reaches are never touched. The heap peak falls to 8.3 MB.
- Applying hunks to a blob keeps slices into the original and into the parsed hunks instead of a `Cow` per line.
- A partial restore hashes the restored content only when the path has an `ident`, filter or line-ending conversion. It refreshes the index stat data only for files whose every hunk was restored, because a file that keeps a hunk can't match the index. The worktree rules are built once.

Rows from `grit-bench daily` on `dist` builds against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0. "base" is `fce803fe8`. Values are medians over three alternating rounds of 30 runs (20 for the second group), in **milliseconds / MB**. Reports: `daily-restore-hunks-6274-before.json` → `daily-restore-hunks-6274-after.json` and `daily-restore-hunks-6274-related-before.json` → `daily-restore-hunks-6274-related-after.json` (first round).

| Scenario | git ms / MB | base ms / MB | after ms / MB |
| --- | ---: | ---: | ---: |
| `restore/hunks@large-file` | 7.05 / 8.6 | 15.18 / 24.3 | **11.22 / 18.4** |
| `add/hunks@large-file` | 12.68 / 10.6 | 17.21 / 22.9 | **14.77 / 17.9** |
| `add/lines@large-file` | 12.93 / 10.7 | 18.08 / 22.9 | **15.83 / 17.7** |
| `log/stat-n50@git-git` | 95.43 / 40.7 | 149.92 / 83.2 | **138.25 / 74.4** |
| `restore/paths@synth-10k` | 11.93 / 12.7 | 11.66 / 16.3 | 11.73 / 16.4 |
| `add/patch-file@large-file` | 12.70 / 10.6 | 8.14 / 14.7 | 8.28 / 14.8 |
| `add/patch-file@synth-10k` | 6.44 / 9.8 | 6.99 / 14.5 | 6.94 / 14.5 |
| `diff/worktree@synth-10k` | 6.72 / 10.0 | 6.33 / 13.2 | 6.70 / 13.2 |
| `diff/worktree@big-blobs` | 315.11 / 190.8 | 209.31 / 193.4 | 204.22 / 193.6 |
| `diff/stat@synth-100k` | 104.60 / 30.3 | 76.02 / 52.1 | 78.14 / 52.2 |
| `show/head@synth-10k` | 1.95 / 8.9 | 2.64 / 11.7 | 2.60 / 11.8 |
| `show/head@big-blobs` | 1.50 / 7.7 | 1.99 / 9.7 | 2.00 / 9.6 |
| `show/head@git-git` | 1.86 / 9.0 | 2.26 / 9.9 | 2.19 / 10.0 |
| `patch-id/commit@git-git` | 2.13 / 13.5 | 2.27 / 9.5 | 2.15 / 9.5 |
| `stash/push-paths@synth-10k` | 72.88 / 20.0 | 29.33 / 20.2 | 29.36 / 20.3 |
| `stash/push-pop@synth-10k` | 198.24 / 18.6 | 118.29 / 18.3 | 119.29 / 18.6 |
| `stash/pop-index@synth-10k` | 61.99 / 18.7 | 35.98 / 22.0 | 36.17 / 22.3 |

The diff rows that moved are within noise, and Git's own times drift just as much between those rounds. Repeated directly (`perf stat`, 200 runs of `grit diff` and 40 of `grit diff --stat`, alternating), the base and the new build measure 5.87 vs 5.87 ms and 76.4 vs 75.9 ms, with identical output. The remaining `restore/hunks` gap is the diff itself (about a quarter of the run), the blob read and index load the hunk ids need, and the roughly 11 MB of binary pages every grit command maps (see [Status memory is code layout](#status-memory-is-code-layout-2026-10-31)).

```bash
./target/dist/grit-bench daily --only restore/hunks@large-file,restore/paths@synth-10k,add/hunks@large-file,add/lines@large-file,add/patch-file@large-file,add/patch-file@synth-10k,diff/worktree@synth-10k,show/head@synth-10k,stash/push-paths@synth-10k,stash/push-pop@synth-10k,stash/pop-index@synth-10k \
  --min-runs 30 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-restore-hunks-6274-after.json
./target/dist/grit-bench daily --only diff/worktree@big-blobs,show/head@big-blobs,diff/stat@synth-100k,show/head@git-git,log/stat-n50@git-git,patch-id/commit@git-git \
  --min-runs 20 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-restore-hunks-6274-related-after.json
```

## The CLI joins the hot text layout (2026-10-31)

`branch/list@many-refs` lists the one local branch of a clone with 10,000 packed refs. The work is tiny: grit runs 1.70 million user-space instructions, Git 1.32 million. Yet grit took 1.86 ms against Git's 1.32 ms and 11.0 MB against 6.7 MB. Both gaps are page faults, 336 against Git's 170, and most of the time a short command spends in the kernel goes to mapping file pages on those faults.

- 101 are copy-on-write faults on `.data.rel.ro`, which the loader relocates at start-up (PIE). Every command pays these, and they stay.
- 77 are code faults, each mapping a 64 KB fault-around window, against Git's 24. Only 19 landed in the library's `.text.grit_hot` section. The other 58 were scattered over `.text`.

Two things scattered them. First, the [library layout](#status-memory-is-code-layout-2026-10-31) leaves out the CLI's own code, and argument parsing (clap building the command tree), dispatch, environment setup and rendering run in every command, about 40 windows of their own. Second, the library's list had already gone stale (`head_file_present` and `FilesRefStore::collect_loose_refs_for_prefix`, run by every ref listing, no longer matched), and it never matched some functions in shipped builds. LLVM merges identical generic instantiations from different crates, and which crate's name the surviving copy keeps (`<String as Display>::fmt` instantiated in `grit`, `serde_json` or `ureq_proto`) differs between the profiled unstripped build and the stripped `dist` one.

What changed:

- `scripts/gen_hot_text_layout.py` now writes two fragments from one callgrind profile. [`grit-lib/hot-text.ld`](https://github.com/gitbutlerapp/grit/blob/main/grit-lib/hot-text.ld) keeps the library's functions. [`grit-cli/hot-text.ld`](https://github.com/gitbutlerapp/grit/blob/main/grit-cli/hot-text.ld) gets everything else the `grit` binary runs (CLI, clap, CLI-instantiated library generics), and `grit-cli/build.rs` links it into `grit` as `.text.grit_cli_hot`, about 2 MB, ahead of `.text`.
- Patterns wildcard a generic's instantiating crate, so merged instantiations match whichever name survives.
- The profile gained branch workloads (`--json`, `--merged`, create, rename, force-delete), and both lists were regenerated.

For `grit branch` on the many-refs clone, code faults drop from 77 to 36 (21 library, 10 CLI, 5 left in `.text`, among them the C start-up code), and total faults from 336 to 299.

Rows from `grit-bench daily` on `dist` builds against the system `git` in one sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0. "base" is `8ccac076d`. Times are medians of all samples from three alternating rounds of 30 runs (the related rows: two rounds of 20, five rounds of 40 for `status/clean@small`), and peak RSS is the median of the rounds' peaks. Values are **milliseconds / MB**. Reports: `daily-branch-cli-hot-text-before.json` → `daily-branch-cli-hot-text-after.json` (first round) and `daily-cli-hot-text-related-before.json` → `daily-cli-hot-text-related-after.json`.

| Scenario | git ms / MB | base ms / MB | + CLI layout ms / MB | RSS vs git | time vs git |
| --- | ---: | ---: | ---: | ---: | ---: |
| `branch/list@many-refs` | 1.32 / 6.7 | 1.86 / 11.0 | 1.74 / **8.8** | 1.65× → 1.32× | 1.40 → 1.32 |
| `branch/list-json@many-refs` | 1.50 / 7.4 | 2.08 / 11.5 | 1.91 / **9.3** | 1.55× → 1.25× | 1.38 → 1.27 |
| `branch/list@many-refs-loose` | 1.30 / 6.7 | 1.89 / 10.9 | 1.77 / **8.9** | 1.62× → 1.33× | 1.46 → 1.36 |
| `branch/merged@many-refs` | 1.44 / 7.4 | 2.07 / 11.6 | 1.90 / **9.7** | 1.58× → 1.31× | 1.44 → 1.32 |
| `branch/create@many-refs` | 1.56 / 7.5 | 2.52 / 12.3 | 2.22 / **9.6** | 1.64× → 1.28× | 1.62 → 1.42 |
| `branch/delete@many-refs` | 4.43 / 7.7 | 3.86 / 12.6 | 3.35 / **10.2** | 1.64× → 1.32× | 0.87 → 0.76 |
| `branch/rename@many-refs` | 1.70 / 7.4 | 3.45 / 10.1 | 3.36 / **9.4** | 1.36× → 1.27× | 2.03 → 1.98 |
| `branch/merged-ref@many-refs` | 0.76 / 6.2 | 1.73 / 8.0 | 1.70 / **7.8** | 1.30× → 1.26× | 2.28 → 2.24 |
| `tag/list@many-refs` | 3.27 / 7.8 | 3.48 / 11.8 | 3.35 / **9.8** | 1.51× → 1.25× | 1.06 → 1.02 |
| `status/clean@small` | 1.83 / 7.2 | 2.73 / 11.6 | 2.53 / **9.3** | 1.61× → 1.29× | 1.49 → 1.39 |
| `status/dirty@small` | 1.85 / 7.2 | 2.76 / 11.9 | 2.48 / **9.4** | 1.65× → 1.30× | 1.49 → 1.34 |
| `status/dirty@synth-10k` | 10.52 / 9.3 | 7.70 / 15.1 | 7.47 / **12.9** | 1.61× → 1.38× | 0.73 → 0.71 |
| `show/head@synth-10k` | 1.96 / 8.3 | 2.87 / 13.4 | 2.61 / **11.0** | 1.62× → 1.33× | 1.46 → 1.33 |
| `diff/worktree@synth-10k` | 6.77 / 9.6 | 6.69 / 14.8 | 6.47 / **12.3** | 1.54× → 1.28× | 0.99 → 0.96 |
| `commit/small@synth-10k` | 26.91 / 13.4 | 16.23 / 19.0 | 15.90 / **15.4** | 1.42× → 1.15× | 0.60 → 0.59 |
| `add/modified@synth-10k` | 16.81 / 12.0 | 10.52 / 18.7 | 9.89 / **16.7** | 1.56× → 1.39× | 0.63 → 0.59 |
| `switch/branch@synth-10k` | 22.07 / 10.3 | 15.87 / 14.7 | 15.36 / **12.5** | 1.43× → 1.22× | 0.72 → 0.70 |
| `reset/mixed@synth-10k` | 8.16 / 9.2 | 10.20 / 17.3 | 10.25 / **13.6** | 1.87× → 1.47× | 1.25 → 1.26 |
| `config/get@small` | 1.21 / 6.2 | 1.66 / 10.0 | 1.58 / **7.6** | 1.61× → 1.23× | 1.37 → 1.30 |
| `notes/list@synth-10k` | 1.25 / 6.2 | 2.03 / 10.4 | 1.76 / **9.1** | 1.68× → 1.47× | 1.62 → 1.41 |
| `worktree/list@synth-10k` | 1.59 / 6.5 | 2.50 / 10.2 | 2.31 / **8.2** | 1.58× → 1.27× | 1.58 → 1.46 |
| `merge/clean@synth-10k` | 98.76 / 13.6 | 52.67 / 18.4 | 51.32 / **16.8** | 1.35× → 1.23× | 0.53 → 0.52 |
| `remote/show@many-refs` | 34.84 / 29.8 | 18.08 / 16.4 | 18.43 / **15.3** | 0.55× → 0.51× | 0.52 → 0.53 |

`branch/merged-ref` and `branch/rename` time `grit-bench` drivers that call the library directly, so only the refreshed library list reaches them. Every one of the 30 related scenarios in the reports uses less memory. `status/clean@synth-10k` reads flat (6.58 → 6.56 ms, 14.9 → 14.6 MB): its peak is the first run after the prepare step, which takes the threaded index refresh, and a settled run directly under GNU `time` drops from 11.4 to 9.2 MB. Time changes for rows over 10 ms are within run-to-run noise.

```bash
./target/dist/grit-bench daily --only branch/list@many-refs,branch/list-json@many-refs,branch/list@many-refs-loose,branch/merged@many-refs,branch/merged-ref@many-refs,branch/create@many-refs,branch/delete@many-refs,branch/rename@many-refs,tag/list@many-refs \
  --min-runs 30 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-branch-cli-hot-text-after.json
```

## Sparse switch keeps changed excluded directories opaque (2026-10-31)

`switch/sparse@sparse-unique-120k` changes 5,000 paths inside the 400 excluded
directories of a sparse-index checkout. A switch expanded every excluded
directory whose subtree differed (about 120,000 index rows), diffed and probed
those rows, then folded them back into sparse-directory entries on write. After
the switch it also reloaded the index fully expanded, which every sparse switch
paid, only to look for case-folding collisions on repositories that ignore case.

A sparse-directory row that still records `HEAD`'s subtree and has no files on
disk holds nothing staged and nothing to protect, so the switch now points it at
the target subtree without expanding it, as Git's two-way merge does for sparse
directories. Rows holding other trees (staged content) are still expanded and
checked. The head-to-target tree diff stops at the opaque rows, the cache-tree is
rebuilt from the remaining valid nodes without reading excluded subtrees, and
the index is reloaded after a switch only when `core.ignorecase` is set.

Daily-matrix medians of ten runs (each run goes out and back), measured in one
sitting on a four-vCPU Intel Xeon Linux VM with stock Git 2.43.0 and Rust 1.99.0
release builds. Values are **milliseconds / peak MiB**:

| Scenario | Git | Before | After |
| --- | ---: | ---: | ---: |
| `switch/sparse@sparse-unique-120k` | 73.3 / 17.7 | 367.0 / 56.6 | 24.0 / 23.0 |
| `switch/sparse@sparse-120k` | 42.1 / 7.8 | 104.9 / 39.6 | 23.0 / 22.2 |
| `switch/sparse@sparse-240k` | 62.0 / 8.1 | 190.4 / 58.1 | 22.8 / 22.2 |
| `reset/sparse@sparse-unique-120k` | 51.2 / 22.0 | 36.6 / 22.0 | 36.9 / 22.8 |
| `reset/sparse@sparse-120k` | 27.2 / 12.1 | 35.7 / 21.4 | 35.4 / 22.2 |
| `reset/sparse@sparse-240k` | 36.6 / 16.9 | 38.2 / 21.4 | 35.9 / 22.2 |
| `switch/branch@synth-100k` | 173.9 / 35.9 | 108.1 / 44.2 | 97.5 / 44.6 |
| `switch/branch@synth-10k` | 21.6 / 10.2 | 19.7 / 22.8 | 16.7 / 23.4 |
| `switch/delete-heavy@wide-48k` | 122.9 / 21.7 | 134.4 / 34.7 | 129.2 / 35.7 |
| `switch/large-blob@big-asset` | 866.4 / 306.9 | 539.4 / 20.2 | 802.4 / 20.5 |

The changed-tree switch is **15×** faster and now **3×** faster than Git; peak
memory drops from 56.6 to 23.0 MiB. The other rows never keep a changed sparse
directory and stay within noise: a second run in reverse order measured
`switch/large-blob@big-asset` at 774.4 ms before and 776.3 ms after, and
`switch/branch@synth-100k` at 105.8 and 85.7 ms. Full reports:
[`daily-switch-sparse-unique-5841-before.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-switch-sparse-unique-5841-before.json)
and
[`daily-switch-sparse-unique-5841-after.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-switch-sparse-unique-5841-after.json).

## Status memory is code layout (2026-10-31)

`grit status` used about twice Git's peak RSS on `synth-10k` (20.1 MB against 9.7 MB on `status/clean@synth-10k`). None of the gap is data. Heaptrack puts the whole run's heap peak at 1.75 MB, and at exit a status run holds 1.35 MB of anonymous memory against Git's 2.0 MB. The index, untracked walk and ignore rules are already smaller than Git's. The rest is file-backed pages of the binary:

- A status run executes about 880 functions, 1.4 MB of machine code, but they were spread over 135 of the 392 64 KB windows of the 25 MB `.text`. The kernel maps file-backed text a whole fault-around window at a time, so about 8.5 MB of code became resident against Git's 1.85 MB. `madvise` cannot turn fault-around off.
- How much of that is fixed cost: `grit --version` took 5.9 MB, `status` on a one-file repository 13.3 MB, and on 10,000 files only 1.3 MB more.
- Every operation has the same problem. Repository discovery, config, refs, object and pack reads and the index are library code shared by every command, scattered like the rest.

Rules that need no profile do not help. Simulating layouts from `nm` sizes and callgrind's executed set:

- Grouping whole crates or the library's core modules leaves status at 138–146 windows, against 147 for the original layout.
- Hot/cold splitting (`-C llvm-args=-hot-cold-split` with `-z keep-text-section-prefix`) moved only 0.85 MB out and made status RSS worse.
- What works is placing the functions that run next to each other.

`grit-lib` now has a `hot-text-layout` feature, and `grit` enables it:

- [`grit-lib/hot-text.ld`](https://github.com/gitbutlerapp/grit/blob/main/grit-lib/hot-text.ld) lists the library's hot code path: functions of `grit-lib`, the Rust runtime and grit-lib's own dependencies that the library's common operations run. Anything naming another crate (a CLI type, clap, output rendering) is left out.
- `grit-lib/build.rs` publishes the fragment as a native library, so it reaches every binary that links grit-lib, embedders included. LLD reads it as an implicit linker script and places the functions in one `.text.grit_hot` section, about 3.2 MB, in front of `.text`.
- Functions are ordered by how many workloads run them, so the core every operation shares comes first.
- The feature acts only on `x86_64-unknown-linux-gnu`, where rustc links with LLD.
- See [Hot text layout](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#hot-text-layout) for how it is generated.

The CLI's own code (argument parsing, dispatch, rendering) keeps its default placement. That is most of what still separates this from a layout that also orders the binary's own functions: in a direct run on `synth-10k` such a layout brought settled `status` to 9.3 MB, while the library layout gives 11.5 MB, from 14.2 MB.

Rows from `grit-bench daily` on `dist` builds, before and after run back to back against the system `git`: status rows with 15 runs and 3 warmups, the others with 10 runs and 2 warmups. The setup was a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0, and "base" is `3c14b849f`. Reports: `daily-status-hot-text-before.json` → `daily-status-hot-text-after.json` and `daily-hot-text-related-before.json` → `daily-hot-text-related-after.json`. Peak RSS is the largest sample of the row, taken through `/bin/sh` (about 2 MB above a direct run for either tool), so it includes the first run after each prepare hook.

| Scenario | git ms / MB | base ms / MB | + layout ms / MB | RSS vs git | time vs git |
| --- | ---: | ---: | ---: | ---: | ---: |
| `status/clean@synth-10k` | 7.31 / 9.7 | 7.86 / 20.1 | 7.17 / **17.6** | 2.07× → 1.82× | 1.05 → 0.98 |
| `status/dirty@synth-10k` | 10.69 / 9.7 | 9.16 / 17.1 | 7.23 / **14.4** | 1.76× → 1.48× | 0.78 → 0.68 |
| `status/renamed@synth-10k` | 7.56 / 9.7 | 7.37 / 19.9 | 7.71 / **16.7** | 2.06× → 1.72× | 1.04 → 1.02 |
| `status/clean@small` | 1.66 / 7.4 | 3.12 / 15.3 | 2.46 / **10.9** | 2.04× → 1.47× | 1.59 → 1.48 |
| `status/dirty@small` | 1.76 / 7.5 | 3.71 / 15.6 | 2.53 / **11.6** | 2.06× → 1.55× | 1.72 → 1.44 |
| `status/clean@synth-100k` | 70.07 / 30.4 | 51.74 / 35.6 | 52.28 / **34.3** | 1.18× → 1.13× | 0.75 → 0.75 |
| `status/dirty@synth-100k` | 103.43 / 30.2 | 61.12 / 39.9 | 54.91 / **37.1** | 1.32× → 1.23× | 0.58 → 0.53 |
| `status/paths@synth-100k` | 48.71 / 30.3 | 53.45 / 40.0 | 45.30 / **36.8** | 1.32× → 1.21× | 1.09 → 0.93 |
| `status/clean@deep-tree` | 1140.87 / 40.6 | 388.16 / 42.8 | 392.21 / **41.4** | 1.06× → 1.02× | 0.34 → 0.34 |
| `status/dirty@deep-tree` | 1231.58 / 41.3 | 453.70 / 54.7 | 439.86 / **51.6** | 1.32× → 1.25× | 0.38 → 0.36 |
| `diff/worktree@synth-10k` | 7.07 / 9.9 | 7.22 / 17.6 | 7.20 / **14.9** | 1.77× → 1.50× | 1.07 → 1.02 |
| `show/head@synth-10k` | 2.00 / 8.9 | 3.12 / 16.7 | 2.82 / **14.1** | 1.95× → 1.59× | 1.54 → 1.41 |
| `add/modified@synth-10k` | 16.81 / 12.4 | 11.75 / 23.0 | 10.68 / **20.1** | 1.85× → 1.61× | 0.69 → 0.64 |
| `commit/small@synth-10k` | 27.77 / 13.7 | 18.12 / 22.1 | 17.12 / **20.2** | 1.59× → 1.47× | 0.64 → 0.62 |
| `switch/branch@synth-10k` | 25.76 / 10.7 | 20.12 / 18.7 | 21.35 / **15.9** | 1.75× → 1.48× | 0.82 → 0.83 |
| `stash/push-pop@synth-10k` | 222.78 / 18.4 | 144.72 / 22.7 | 149.13 / **19.7** | 1.23× → 1.07× | 0.60 → 0.67 |
| `reset/hard@synth-10k` | 16.00 / 11.1 | 21.82 / 19.4 | 22.67 / **17.6** | 1.75× → 1.59× | 1.41 → 1.42 |
| `merge/clean@synth-10k` | 126.07 / 14.3 | 57.27 / 22.0 | 63.72 / **19.5** | 1.54× → 1.37× | 0.53 → 0.51 |
| `notes/list@synth-10k` | 1.37 / 7.0 | 2.02 / 13.0 | 1.97 / **11.6** | 2.06× → 1.66× | 1.50 → 1.44 |
| `reflog/list@synth-10k` | 3.90 / 9.5 | 4.43 / 12.9 | 4.37 / **12.2** | 1.35× → 1.29× | 1.16 → 1.12 |
| `clean/untracked@synth-10k` | 4.68 / 8.7 | 5.07 / 14.8 | 4.94 / **13.4** | 1.65× → 1.53× | 1.15 → 1.06 |
| `gc/pack-refs@synth-10k` | 1.70 / 7.1 | 3.59 / 12.7 | 3.19 / **12.0** | 1.82× → 1.70× | 2.10 → 1.88 |

`notes`, `reset`, `merge`, `clean` and `gc` are not in the profiling workload; they shrink because they share the library core.

Single-sitting times drift with the machine. Alternating rounds show no slowdown:

| Command | base ms | + layout ms |
| --- | ---: | ---: |
| `status` on `synth-10k`, dirty, prepare before every run (5 × 20) | 7.78 | **7.46** |
| `status` on `synth-10k`, clean, prepare before every run (5 × 20) | 7.28 | **6.98** |
| `status` on `synth-100k`, dirty, prepare before every run (5 × 8) | 58.66 | 55.73 |
| `status` on `small`, settled (5 × 100) | 2.35 | **2.17** |
| `diff` on `synth-10k`, settled (4 × 60) | 5.50 | **5.06** |
| `log --limit 20` on `synth-10k` (4 × 60) | 2.19 | 2.10 |
| `merge/clean@synth-10k`, three alternating daily runs | 55.21 | 54.55 |
| `stash/push-pop@synth-10k`, three alternating daily runs | 128.82 | 126.79 |
| `status/renamed@synth-10k`, three alternating daily runs | 7.22 | 6.76 |

GNU `time` peak RSS on `synth-10k` (median of three):

- Status: settled 14.2 → 11.5 MB.
- Diffs and history: `diff` 14.6 → 12.3 MB, `log --limit 20` 12.2 → 10.7 MB, `show HEAD` 13.7 → 11.9 MB, `blame` 19.2 → 17.5 MB.
- Refs: `branch` 10.3 → 9.3 MB, `tag` 10.3 → 9.2 MB.

What remains of the status gap is the CLI's own scattered code plus read-only data: `.rela.dyn` (576 KB, read whole by the loader) and the `.rodata` and `.data.rel.ro` pages startup touches, about 2.2 MB.

```bash
./target/dist/grit-bench daily --only status/clean@synth-10k,status/dirty@synth-10k,status/renamed@synth-10k,status/clean@synth-100k,status/dirty@synth-100k,status/paths@synth-100k,status/clean@deep-tree,status/dirty@deep-tree,status/clean@small,status/dirty@small \
  --min-runs 15 --warmup 3 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-status-hot-text-after.json
./target/dist/grit-bench daily --only diff/worktree@synth-10k,show/head@synth-10k,add/modified@synth-10k,commit/small@synth-10k,switch/branch@synth-10k,stash/push-pop@synth-10k,reset/hard@synth-10k,notes/list@synth-10k,reflog/list@synth-10k,merge/clean@synth-10k,clean/untracked@synth-10k,gc/pack-refs@synth-10k \
  --min-runs 10 --warmup 2 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-hot-text-related-after.json
```
## Protocol v0 fetches over smart HTTP negotiate like Git (2026-11-01)

Since fetch honours `protocol.version` (2026-10-31), the daily matrix's `git http-backend` scenarios run grit on protocol v0, because the bench's global config sets `protocol.version=0`. That path had three problems:

- **Lost requests.** Git's stateless upload-pack reads a v0 request only up to the first flush after its `have` lines, then exits. grit's HTTP connection replayed every earlier have round, flushes included, so the server answered the first round again and never saw `done`. Any fetch that sent haves got no pack, and `fetch/incremental`, `fetch/one-branch`, `pull/check` and `pull/fast-forward` on git.git failed their correctness checks. `fetch/dry-run` and `fetch/incremental-http` passed but took over 700 ms.
- **A request line Git rejects.** A v0 `--deepen` sent `deepen-relative` as a request line as well as a capability. Git's upload-pack accepts it only as a capability, so it hung up on the request: `fetch/deepen` failed with "unexpected EOF after done".
- **No haves when shallow.** A v0 shallow request offered no haves, so a deepen of git.git received 59k objects (43 MiB), as v2 did before the [earlier fix](#deepening-a-shallow-fetch-negotiates-2026-10-29).

`grit-lib` now frames stateless v0 requests the way Git's `fetch-pack` does. Each POST is one complete request: the want block, then every `have` the server has acknowledged as `common`, then the next batch of haves and a flush, or `done`. The connection sends each body as written instead of replaying earlier rounds. Other changes:

- A request without a deepen sends its want block with the first have round, not in a POST of its own.
- Negotiation stops once the server says `ready`.
- With a server that offers `no-done`, the pack follows `ready` directly, without one more POST that only carries `done`.
- Rounds grow geometrically, as in v2.
- The ACKs after `done` are read up to the final bare `ACK`.

Shallow requests negotiate as in v2: the tips being deepened are offered as haves, no advertised tip is marked common in advance (Git skips that when deepening too), and the shallow list that every stateless reply repeats is consumed before its ACKs. Tests in `grit-lib/tests/fetch_v0_http_git_compat.rs` fetch from the system `git http-backend` (a new `grit_test_support::git_http_backend` server) on v0 and v2. They check each v0 request body, the shallow boundaries against git's, the object counts, and `git fsck`. They fail on the old code.

Rows from `grit-bench daily` (10 runs, 3 warmups, `release` builds) against the system `git` in the same sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0. "daily build" is `8dbc58e77` (the 2026-10-27 census build, which still spoke v2), "main" is `7d41ce6df`. "fail" means grit failed the scenario's correctness check. Values are **milliseconds / MB**. Reports: `daily-fetch-v0-6196-daily-build.json`, `daily-fetch-v0-6196-before.json` → `daily-fetch-v0-6196-after.json`.

| Scenario | git | daily build | main | + change |
| --- | ---: | ---: | ---: | ---: |
| `fetch/deepen@git-git` | 665.1 / 270.9 | 5425.4 / 78.1 | fail | **173.7** / 60.8 |
| `fetch/incremental@git-git` | 111.9 / 70.9 | 47.5 / 50.4 | fail | **55.6** / 52.3 |
| `fetch/one-branch@git-git` | 105.8 / 71.0 | 50.2 / 50.3 | fail | **53.6** / 52.4 |
| `fetch/dry-run@git-git` | 104.6 / 70.7 | 47.4 / 50.9 | 721.9 / 69.2 | **52.0** / 52.7 |
| `fetch/incremental-http@git-git` | 153.6 / 70.8 | 90.2 / 50.1 | 740.2 / 69.4 | **91.7** / 52.2 |
| `pull/check@git-git` | 117.5 / 71.0 | 48.3 / 50.5 | fail | **61.3** / 52.4 |
| `pull/fast-forward@git-git` | 117.4 / 76.4 | 56.4 / 50.4 | fail | **58.0** / 52.4 |
| `fetch/noop@git-git` | 24.8 / 56.3 | 11.9 / 19.1 | 8.6 / 20.2 | 8.8 / 20.4 |
| `fetch/noop-local@many-refs` | 20.7 / 23.6 | 19.1 / 19.9 | 15.5 / 20.7 | 15.4 / 21.6 |
| `fetch/noop-local-no-tags@many-refs` | 19.2 / 23.8 | 14.2 / 19.8 | 11.7 / 20.7 | 11.9 / 21.4 |
| `clone/shallow@git-git` | 1592.1 / 39.8 | 1658.0 / 57.4 | 1623.7 / 58.0 | 1623.8 / 56.1 |
| `clone/blobless@git-git` | 8683.5 / 73.5 | 7171.8 / 66.9 | 7575.9 / 149.0 | 7299.1 / 149.8 |
| `clone/http@synth-10k` | 758.6 / 36.9 | 697.4 / 33.6 | 602.6 / 50.5 | 606.0 / 50.2 |
| `clone/shallow-depth1-http@synth-10k` | 191.9 / 32.1 | 198.5 / 31.2 | 200.5 / 31.9 | 202.2 / 32.5 |
| `ls-remote/all@many-refs` | 14.5 / 28.0 | 14.4 / 18.8 | 13.5 / 19.7 | 13.6 / 20.1 |
| `push/incremental-http@git-git` | 1641.8 / 333.7 | 959.3 / 90.7 | 962.3 / 92.2 | 979.4 / 92.8 |

The clone, `ls-remote` and push rows send the same requests as before; three alternating reruns of `clone/shallow-depth1-http` (main 190–195 ms, this change 194–199 ms, git 194–229 ms) and `push/incremental-http` (974–987 ms against 978–991 ms) are within run-to-run noise. The incremental rows remain a few milliseconds above the daily build, which spoke protocol v2: a v0 `info/refs` advertises every ref (955 on git.git), and v2 asks only for the refs a refspec can match.

```bash
./target/release/grit-bench daily --only fetch/deepen@git-git --grit ./target/release/grit \
  --min-runs 10 --warmup 3 --format json --output /tmp/fetch-deepen.json
```

## Deepening a shallow fetch negotiates (2026-10-29)

A fetch with `--deepen`, `--depth` or into a shallow repository used to send the server no `have` lines, so the server packed every object of the deepened history. On `fetch/deepen@git-git` (a depth-1 `master` deepened by 50 commits, served by `git http-backend`), grit received 59k objects (43 MiB) that the client already held through its other branches, and spent most of its 5.5 s waiting for the server to build that pack and then indexing it. Git negotiates and receives 175 objects.

`grit-lib` now negotiates shallow fetches the way Git does. The skipping negotiator already treats local shallow commits as parentless, and the server limits each `have` to the client's side of its `shallow` lines. The tips being deepened are offered as haves too. On the serving side, `grit upload-pack` no longer uses reachability bitmaps when the client is shallow, because a bitmap closure of a `have` runs past the client's boundary. Before this change, a system `git fetch --unshallow` (protocol v2) from a bitmapped repository failed with "remote did not send all necessary objects". Git's `pack-objects` also turns bitmaps off for shallow clients.

Rows from `grit-bench daily` (10 runs, 2 warmups, `release` builds) against the system `git` in the same sitting, on a four-vCPU Intel Xeon VM with Git 2.43.0. "base" is `e97f684a5`. Reports: `daily-fetch-deepen-5548-before.json` → `daily-fetch-deepen-5548-after.json`.

| Scenario | git ms / MB | base | + change |
| --- | ---: | ---: | ---: |
| `fetch/deepen@git-git` | 656.6 / 271.4 | 5498.0 / 81.5 | **119.3** / 57.4 |
| `fetch/one-branch@git-git` | 106.2 / 70.9 | 48.3 / 51.5 | 47.9 / 51.4 |
| `fetch/incremental@git-git` | 106.0 / 70.9 | 47.1 / 51.5 | 47.7 / 51.2 |
| `fetch/incremental-http@git-git` | 126.8 / 71.1 | 51.1 / 51.4 | 49.7 / 50.9 |
| `fetch/dry-run@git-git` | 105.7 / 70.7 | 46.9 / 51.8 | 46.7 / 52.0 |
| `fetch/noop@git-git` | 25.2 / 56.3 | 10.7 / 20.0 | 11.4 / 20.0 |
| `fetch/noop-local@many-refs` | 20.7 / 23.6 | 20.6 / 20.9 | 16.6 / 21.1 |
| `fetch/noop-local-no-tags@many-refs` | 19.2 / 23.8 | 15.3 / 20.9 | 12.8 / 20.8 |
| `clone/shallow@git-git` | 1579.7 / 40.0 | 1671.7 / 57.7 | 1686.4 / 56.5 |
| `clone/shallow-depth1@synth-10k` | 180.2 / 32.2 | 183.3 / 32.2 | 181.6 / 32.4 |
| `clone/shallow-depth1-http@synth-10k` | 185.2 / 32.2 | 188.5 / 32.3 | 189.0 / 31.7 |

Only the deepen row changes. A shallow clone has no local refs to offer, and the other fetches were not shallow requests. The protocol v0 fetch path still skips haves for shallow requests.

## Log pages: one config parse, one pack reader (2026-10-29)

`grit log --limit 100` on git.git spends about as long per commit as Git (grit's
inflate is faster than system zlib), so the gap is fixed cost. Four pieces of
that cost were repeated work:

- Opening a repository parsed the config cascade twice. `Odb::hash_algo` runs
  while the ref store is set up and loaded the cascade under its own cache key,
  so the repository's config session missed it and parsed every file again.
  Both now share one entry, and the cache hands out its `Arc<ConfigSet>`
  instead of deep-copying every entry for each reader. Every command that
  opens a repository saves about 0.1 ms.
- The log page walk never ran under the object database's pack-read context,
  so its single-pack reader stayed closed and each commit read installed the
  context and opened a reader again (250 reader opens for 100 commits). The
  walk now reads through one reader. Objects that reader does not hold
  (loose ones, or ones in an alternate) are read through the object database,
  so they never look like missing parents.
- Abbreviating commit ids after the walk listed the pack indexes outside that
  context, so `grit log --oneline` mapped git.git's 10 MB `.idx` a second time
  (4.5 MB more resident, 0.2 ms).
- Plain human and Markdown rows built a JSON row per commit (hex id, two parsed
  identities with RFC 3339 dates, a relative date) that only `--json` and
  `--decorate` read.

Measured on a 4-vCPU Intel Xeon VM, system Git 2.43.0, `dist` builds of main
(`1e65b2ed2`) and of the change, in one sitting of four `grit-bench daily`
runs ordered main, change, change, main (15 timed runs after 5 warmups each).
Times average the two medians per binary. Peak RSS is the lower of each
binary's two single samples. Output of `grit log` is byte-identical in human,
Markdown and `--json` modes, with and without `--graph`, `--decorate`,
`--stat`, `-p`, paths and filters.

| Scenario | Git ms | Main ms | Change ms | Main KiB | Change KiB |
| --- | ---: | ---: | ---: | ---: | ---: |
| `log/oneline-100@git-git` | 2.92 | 3.59 | 3.52 | 19,584 | 20,192 |
| `log/oneline-100@synth-100k` | 2.22 | 2.98 | 2.72 | 15,860 | 16,404 |
| `log/all@git-git` | 45.10 | 53.04 | 44.64 | 34,548 | 33,388 |
| `log/json-100@git-git` | 2.98 | 3.98 | 3.94 | 20,808 | 21,276 |
| `log/author@git-git` | 2.33 | 3.68 | 3.65 | 18,632 | 19,400 |
| `log/path@git-git` | 343.28 | 359.87 | 353.10 | 273,700 | 274,136 |
| `shortlog/ahead@git-git` | 16.16 | 2.15 | 2.08 | 11,808 | 12,000 |
| `show/head@git-git` | 1.91 | 2.63 | 2.56 | 13,552 | 14,048 |

`log/stat-n50`, `log/author-n50@deep-history-cg`, `log/json-filter`,
`log/limit-path` and `log/json-n10-loose-tags` came out 1–3% slower in this
sitting. Separate interleaved checks of the same commands (40–200 runs per
binary, alternating) had the change level or 1–4% ahead on every one of them.
Interleaved runs of `log --limit 100` alone measure 3.50 ms on main and 3.23 ms
with the change.

Peak RSS reads about 0.5 MB higher with the change on most rows, including
`status`, which none of these changes touch. The heap is 100 KB smaller and
`.text` is 7 KB smaller. What moved is which 64 KB fault-around windows of
the 24 MB text segment the run touches: rebuilding under fat LTO reshuffles
function placement (see
[`show` HEAD is startup-bound](#show-head-is-startup-bound-2026-10-28)). The
only RSS this change adds by design is two more `.idx` windows (128 KB) that
the single-pack reader's ordered lookups touch. Raw samples:
[`daily-log-startup-5443-before-1.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-log-startup-5443-before-1.json),
[`daily-log-startup-5443-after-1.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-log-startup-5443-after-1.json),
[`daily-log-startup-5443-after-2.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-log-startup-5443-after-2.json)
and
[`daily-log-startup-5443-before-2.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-log-startup-5443-before-2.json).

What remains of the gap on `log/oneline-100@git-git` is process-wide: about
110 PIE relocation faults, code pages spread over the text segment (6.9 MB
resident against Git's 1.8 MB), and clap copying every global flag into each
of the 55 subcommands before parsing (270 KB of the 375 KB heap peak).

## Clean status: tracked files leave the untracked walk early (2026-10-29)

Most entries in a work-tree directory are tracked files, and the untracked walk has nothing to report for them. It still copied each name, built the file's path, looked it up in a hash set of every index path, and sorted the whole listing. That set took one insert per entry and was built serially before the walk could start. The walk also opened every directory's `.gitignore` as it entered the directory. On `status/clean@synth-10k` the walk used about 6 ms of CPU across four threads, where Git's single-threaded walk uses about 2.3 ms.

Three changes in `grit-lib`:

- Each directory's names are matched against that directory's run of the sorted index while it is read. Tracked files are dropped immediately, and only the remaining entries are kept and sorted. Exact-case tracked lookups binary-search the sorted index instead of building the set.
- A directory's `.gitignore` is read on the first ignore check for a path inside it. As in Git, a directory whose entries are all tracked never opens it. Results do not change, because ignore checks already loaded missing parent rules on demand.
- The worktree diff no longer allocates a slot per index entry for stat-refresh content checks when no entry needs one. On 10k entries that was about 400 KB written on the serial path.

On `synth-10k` the untracked walk dropped from about 2.1 ms to 1.3 ms of wall time (6.0 ms to 3.6 ms of CPU), and the stat-refresh step dropped from 0.16 ms to 0.05 ms.

The status rows below come from `grit-bench daily` (20 runs, 5 warmups) on `dist` builds, measured against the system `git` in the same sitting. The setup was a four-vCPU Intel Xeon VM with Git 2.43.0 and Rust 1.99.0, and "base" is `e6e6f6c21`. Reports: `daily-perf-5414-status-before.json` → `daily-perf-5414-status-after.json`.

| Scenario | git ms / MB | dist base | dist + change |
| --- | ---: | ---: | ---: |
| `status/clean@synth-10k` | 7.22 / 9.2 | 8.83 / 16.2 | **7.74** / 15.3 |
| `status/dirty@synth-10k` | 11.40 / 9.4 | 9.57 / 17.8 | **8.39** / 18.4 |
| `status/renamed@synth-10k` | 7.03 / 9.4 | 8.85 / 18.9 | **7.12** / 15.4 |
| `status/clean@synth-100k` | 68.20 / 29.0 | 67.04 / 34.0 | **50.33** / 34.1 |
| `status/dirty@synth-100k` | 99.31 / 29.0 | 61.69 / 39.1 | **50.47** / 37.7 |
| `status/paths@synth-100k` | 44.51 / 29.0 | 47.86 / 39.3 | **42.21** / 37.8 |
| `status/clean@deep-tree` | 1103.34 / 38.7 | 534.49 / 40.9 | **375.27** / 41.1 |
| `status/dirty@deep-tree` | 1191.14 / 39.5 | 586.60 / 52.6 | **423.87** / 51.6 |
| `status/clean@small` | 1.84 / 7.2 | 3.03 / 13.9 | 3.08 / 14.8 |
| `status/dirty@small` | 1.84 / 7.2 | 3.01 / 14.4 | 3.04 / 14.8 |

The `small` rows (100 files) are within noise: five alternating hyperfine rounds of 150 runs put base and change at 2.59 and 2.53 ms on `dist` (12.5 and 12.4 MB). For the headline row, five alternating rounds of 120 runs (median of medians) give:

| `status/clean@synth-10k` | ms | peak RSS MB |
| --- | ---: | ---: |
| git | 6.67 | 7.0 |
| `dist` base → change | 7.86 → **6.71** | 14.0 → 13.9 |
| `release` base → change | 8.06 → **7.26** | 18.4 → 18.1 |

The memory gap is the binary, not data. At exit, a `release` status on `synth-10k` holds about 1.6 MB of anonymous memory and about 16.7 MB of file-backed pages, most of them instruction pages mapped by 64 KB fault-around (see [`show` HEAD is startup-bound](#show-head-is-startup-bound-2026-10-28)).

Commands that share the walk or the worktree diff, from `daily-perf-5414-related-before.json` → `daily-perf-5414-related-after.json` (15 runs each):

| Scenario | git ms | dist base | dist + change |
| --- | ---: | ---: | ---: |
| `clean/untracked@synth-10k` | 4.39 | 5.61 | **5.07** |
| `clean/only-ignored@synth-10k` | 4.26 | 5.50 | **4.89** |
| `commit/small@synth-10k` | 27.32 | 18.89 | **17.37** |
| `add/all@synth-100k` | 145.23 | 654.97 | **615.87** |
| `add/update@synth-100k` | 99.08 | 90.27 | 89.41 |
| `stash/push-pop@synth-10k` | 204.66 | 126.23 | 126.43 |
| `diff/worktree@synth-100k` | 117.32 | 92.55 | 93.31 |

Five single-sitting rows in that report read 4–11% slower: `add/dry-run`, `add/modified` and `commit/dry-run` on `synth-10k`, `diff/worktree@synth-10k`, and `commit/small@small`. Alternating hyperfine rounds on the same fixtures show no regression:

- `add --dry-run`: 7.66 → 6.94 ms.
- `commit --dry-run`: 24.47 → 23.85 ms.
- `diff`: 6.95 → 6.89 ms.
- Staging the modified files: 9.97 → 9.88 ms.
- `commit` on `small`, with base and change taking turns going first: 21.81 → 21.51 ms.

```bash
cargo build --profile dist -p grit-cli
./target/release/grit-bench daily --only status/clean@synth-10k,status/dirty@synth-10k,status/renamed@synth-10k,status/clean@synth-100k,status/dirty@synth-100k,status/paths@synth-100k,status/clean@deep-tree,status/dirty@deep-tree,status/clean@small,status/dirty@small \
  --min-runs 20 --warmup 5 --grit ./target/dist/grit --format json --output grit-utils/baselines/daily-perf-5414-status-after.json
```

## `show` HEAD is startup-bound (2026-10-28)

Tree diffs now walk entries borrowed from the raw tree payload instead of parsing each tree into owned entries. The ODB session cache keeps either form under one byte budget, so `show` on a one-file commit in git.git's 530-entry root tree makes 4,232 allocations instead of 5,304, with a 403 KB heap peak instead of 463 KB.

Median of three hyperfine `-N` medians (400 runs each), peak RSS the median of eleven `/usr/bin/time` runs, all in one sitting. The setup was a four-vCPU Intel Xeon VM with system Git 2.43.0 and Rust 1.99.0. "Base" is `f6dc0e06d`.

| `show` HEAD | git ms / MB | release base | release + change | dist base | dist + change |
| --- | ---: | ---: | ---: | ---: | ---: |
| `@git-git` | 1.99 / 6.9 | 3.08 / 16.1 | 3.12 / 16.9 | 2.70 / 13.1 | 2.63 / 12.9 |
| `@big-blobs` | 1.47 / 5.2 | 2.82 / 15.0 | 2.88 / 16.0 | 2.43 / 11.9 | 2.38 / 11.8 |
| `@synth-10k` | 1.96 / 6.6 | 3.87 / 16.8 | 3.92 / 18.0 | 3.37 / 13.8 | 3.19 / 13.5 |

On `dist` the change is 2–5% faster with slightly lower RSS. On `release` it reads about 40 µs slower and up to 1 MB larger. That comes from code layout, not work: the release binary takes 176 instruction page faults instead of 166, spread over unrelated code (argument parsing, config, environment setup), because editing `grit-lib` reshuffles its codegen units.

What remains of the gap to Git is process startup. On `dist` it is bounded by the PIE policy ([PIE startup relocation probe](#pie-startup-relocation-probe)), so later speed work should not look for it in `show`'s code:

- Page faults are about 40% of grit's run time: about 420 per run on `dist` (600 on `release`) against Git's 253.
- Applying PIE relative relocations copy-on-write faults every `.data.rel.ro` page at each start. That costs about 0.4 ms on `release`; linking without PIE would remove it, but release builds stay PIE with ASLR.
- Most of the remaining faults are instruction fetches across the `.text` segment, each mapping a 64 KB fault-around window. They make up most of the peak RSS gap, and LTO (`dist`) is what shrinks them.
- Syscall counts match Git (272 against 300), and the work inside the library is about 0.7 ms with no single cost above about 60 µs.

## Bulk add and first commit of a large new tree (2026-10-28)

Issue [#2169](https://github.com/gitbutlerapp/grit/issues/2169): `grit add` of a
large new tree into an empty index peaked at about 1.4 KiB per file, 4.6 times
`git add -A`. Staging built a read request (attributes, metadata, absolute path)
for every untracked file, then a hashed result for every file, before it created
any index entry. It also cloned every path to look for embedded repositories,
and recorded a second copy of every path to skip the racy-entry re-read at index
write. Now it reads, hashes and stores 4,096 files at a time and turns each chunk
straight into compact index entries. Freshly built entries carry an in-memory
"just hashed" bit instead of the extra path copy. The first `grit commit` of such
a tree also lists its files from the index instead of diffing the new tree
against nothing.

The tree has 300,000 three-line files, 100 per leaf directory
(`top{d%30}/mid{(d/30)%10}/pkg{d/300}/leaf{d}`). Rows are one sitting on a
4-vCPU Intel Xeon Linux VM (15 GiB RAM, Rust 1.99 `release`, Git 2.43.0), with
peak RSS and wall time from `/usr/bin/time`. Main is `307186978`. "Objects
present" stages the same files again after `git rm -r --cached .`. Git's first
commit row is `git add -A` followed by `git commit`.

| Operation | Git | Main | Fixed |
| --- | ---: | ---: | ---: |
| add, fresh repository | 25.32 s / 78,840 KB | 7.70 s / 433,272 KB | 7.04 s / 133,120 KB |
| add `--threads 1`, fresh repository | | 16.54 s / 430,668 KB | 16.04 s / 130,388 KB |
| add, objects present | 2.21 s / 87,100 KB | 7.78 s / 435,488 KB | 4.00 s / 131,260 KB |
| first commit (stage all) | 11.23 + 1.40 s / 79,156 KB | 8.50 s / 434,356 KB | 8.12 s / 153,292 KB |

Peak memory drops by 69% for add and 65% for the first commit. Git still uses
less: the remaining peak is mostly the index entries themselves, the staged path
list the command reports, and the serialized index buffer. The
`add-initial-monorepo-20000` baseline (20,000 files) goes from 48,800 KiB to
32,412 KiB with unchanged time. The `add_bulk_memory` integration test checks
grit's peak against `git add -A` on 60,000 files: 5.2 times before, 1.9 times now.

## Walk reads and JSON-only log rows (2026-10-28)

Commit walks that parse objects themselves (log pages, merge bases, ahead/behind,
containment) used to read each commit through `Odb::read`, which allocates a fresh
object and clones it into the object database's session cache. The walk keeps the
parsed linkage and, for `log`, its own copy of the payload, so the clone was never
read again. Walks now inflate every commit into one reused buffer, keeping delta
bases cached as before. `grit log --json` also stops building the human summary
rows (and re-formatting every author header) that JSON output never renders.

Measured on a 4-vCPU Intel Xeon VM, system Git 2.43.0, `dist` builds, 15 runs after
5 warmups per binary in one sitting. Main is `2e65faf26`; output is byte-identical
for `log` in human, Markdown, `--json` and `--json --filter` modes.

| Scenario | Git ms | Main ms | Fixed ms | Main KiB | Fixed KiB |
| --- | ---: | ---: | ---: | ---: | ---: |
| `log/json-100@git-git` | 2.86 | 4.15 | 3.98 | 21,352 | 21,088 |
| `log/all@git-git` | 45.27 | 55.68 | 53.03 | 35,272 | 35,036 |
| `log/author@git-git` | 2.30 | 3.77 | 3.57 | 19,808 | 19,400 |
| `log/json-filter@git-git` | 2.24 | 4.21 | 4.08 | 20,212 | 20,184 |
| `log/oneline-100@git-git` | 2.85 | 3.76 | 3.67 | 19,544 | 20,024 |
| `log/stat-n50@git-git` | 94.90 | 161.12 | 154.20 | 76,760 | 78,148 |
| `merge-base/branches@git-git` | 15.25 | 3.71 | 3.84 | 19,444 | 19,180 |

The few rows that came out slower by up to 3.5% (`merge-base/branches`,
`shortlog/behind`, `log/limit-path`, `log/oneline-100@synth-100k`, `revs/count`)
did not reproduce in separate Hyperfine checks on the same fixtures (60–300 runs;
2 × 20 interleaved runs for `revs/count`): the fixed binary was level or ahead.
`revs` walks through `rev_list`, which this change does not touch. Most of what
remains on `log/json-100` is per-process: the 100
commits inflate faster than Git's zlib does, while the binary's scattered code
keeps about 7 MB resident through fault-around (see
[Tree metadata diff through the CLI](#tree-metadata-diff-through-the-cli-2026-10-27))
and position-independent relocations cost another 0.2 ms at startup. Raw samples:
[`daily-log-walk-4978-before.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-log-walk-4978-before.json)
and
[`daily-log-walk-4978-after.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-log-walk-4978-after.json).

## Range counts on merge-heavy history (2026-10-27)

`grit revs --count A..B` (and `RevWalk::count`) used to decide, commit by
commit, whether each commit in the range was hidden, with a fresh ancestry walk
per merged commit. Counting a range is now one walk that marks the excluded side
down to the common frontier, with `--since`, `--until`, `--merges`,
`--no-merges` and `--first-parent` applied while counting. This was the path for
every count without a commit-graph and for every filtered count with one.

Measured on rust-lang/cargo (23,585 commits) and a bare clone of git/git, with
system Git 2.43.0 on Ubuntu, four vCPUs, release build. Each cell is the median
of three runs, with peak RSS from `/usr/bin/time`. Main is `6be9ffe0a`.

| Count | Graph | Git (s / MB) | Main (s / MB) | Fixed (s / MB) |
| --- | --- | ---: | ---: | ---: |
| cargo `0.50.0..HEAD` (13,489) | none | 0.13 / 25 | 18.60 / 37 | 0.13 / 36 |
| cargo `0.30.0..0.60.0` (5,280) | none | 0.03 / 13 | 0.38 / 25 | 0.03 / 25 |
| cargo `--since 2023-01-01 0.50.0..HEAD` | none | 0.07 / 20 | 18.80 / 36 | 0.14 / 36 |
| cargo `--merges 0.50.0..HEAD` | none | 0.12 / 25 | 18.91 / 37 | 0.13 / 36 |
| cargo `--first-parent 0.50.0..HEAD` | none | 0.03 / 19 | 5.81 / 30 | 0.13 / 36 |
| cargo `--left-right 0.50.0...HEAD` | none | 0.14 / 40 | 19.25 / 40 | 0.16 / 40 |
| git/git `v2.30.0..v2.56.0` (20,852) | none | 0.35 / 51 | 7.57 / 82 | 0.34 / 63 |
| git/git `--since 2024-01-01 v2.30.0..v2.56.0` | none | 0.22 / 36 | 7.53 / 82 | 0.33 / 63 |
| cargo `--since 2023-01-01 0.50.0..HEAD` | Git-written | 0.00 / 13 | 17.64 / 32 | 0.03 / 22 |
| cargo `--merges 0.50.0..HEAD` | Git-written | 0.01 / 14 | 17.62 / 32 | 0.03 / 22 |
| git/git `--since 2024-01-01 v2.30.0..v2.56.0` | Git-written | 0.02 / 24 | 6.90 / 52 | 0.06 / 33 |
| git/git `--first-parent v2.30.0..v2.56.0` | Git-written | 0.00 / 21 | 0.02 / 19 | 0.00 / 20 |

Unfiltered counts with a commit-graph, full-history counts and range listings
stayed within noise of main (cargo `0.50.0..HEAD` listing: 149 ms on both over
15 runs). Graph-backed `--first-parent` counts with an excluded tip also used to
follow only first parents from the excluded side, so they could include commits
that tip reaches through a merge's second parent; they now match Git.

## Bounded revision counts with a commit-graph (2026-10-27)

Unfiltered two-dot counts now reuse the generation-aware hide walk instead of
building both complete ancestor sets. On git/git at `583e98cd091c` (82,537
commits at the topic tip; 56 commits in `HEAD~20..HEAD`), graph-backed counts
are about 20× faster and within 1.6× system Git. The commit-graph remains mapped;
only the range frontier is visited.

Ubuntu 24.04, four vCPUs, Rust 1.99.0, `dist` profile, system Git 2.55.0.
Hyperfine 1.19.0 used `-N`, five warmups and 30 runs per command. Benchmark
processes used a separate physical core from remaining harness tests. RSS is
the median of five process peaks measured with `/usr/bin/time -v`.

| Graph | Git count (ms / MiB) | Main `1c4fdeca7` (ms / MiB) | Fixed (ms / MiB) |
| --- | ---: | ---: | ---: |
| None | 6.76 / 15.75 | 22.60 / 22.56 | 23.34 / 22.48 |
| Git-written | 2.42 / 10.88 | 72.76 / 15.67 | 3.68 / 14.57 |
| grit-written | 2.43 / 10.97 | 74.93 / 15.76 | 3.72 / 14.77 |

Range listings and full-`HEAD` counts were also measured in all three setups.
Their mean times stayed within 3.3% of main, with peak RSS within 0.25 MiB.
The full measurements and reproduction commands are in
[`revs-graph-range-1669.md`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/revs-graph-range-1669.md).
The daily matrix now includes `revs/count-range-commit-graph@git-git`, with
committed before/after reports on its pinned `v2.47.0` fixture.

## Sirocco packed-read evaluation (2026-10-26)

The reference candidate [8a0f13f4](https://but.dev/c/01m5x3zw0zc7c88a4wzq0bjfks/b/factory%2Fsirocco-inflate?commit=8a0f13f4eb4cbc38642313001ca1b22de0567b8d) evaluated Sirocco 0.3.1 in the shared packed-object scratch for full-input reads and prefix probes. Complete reads had to finish a checksum-validated zlib member and match its declared output size, including empty objects. Loose-object reads, bounded streaming, checksum-validating scans, and compression retained flate2/zlib-rs.

**Sirocco was not adopted.** Packed ODB reads improved, but median peak RSS grew by **56–364 KiB** and the **10k packed-history walk without a commit-graph regressed by about 5%**. This misses the requirement for measured wins without compatibility or memory regressions. Grit retains **flate2/zlib-rs as its only decoder** and ships no Sirocco dependency. The tables below document the rejected experiment; they are not published performance baselines for the current implementation.

The baseline is `origin/main` at `99332d04950da7af39beff55b9874c42d17e0d5a`. Frozen before/after `dist` binaries and system Git 2.55.0 ran on the same 4-vCPU AMD EPYC 9V45 VM (16 GiB RAM, Linux, Rust 1.99). Builds and tests were stopped during measurement. The table uses two warmups and 21 fresh-process samples per tool, rotating tool order and pinning all children to CPU 0. Wall time includes the `/usr/bin/time -v` wrapper; RSS is the median of its per-process peak readings. Pinned and unpinned raw samples are preserved in the reference candidate commit.

The synthetic repository has 50,000 commits and 20,000 files in Git-produced packs, with separate commit-graph and no-graph copies. The real repository is this checkout's `main` (11275 commits, limit 5,000). Library walks use `grit-bench lib-op rev-list-head`; the Git equivalent is `git rev-list HEAD`. CLI rows use `log --oneline -n <limit>` with output discarded.

| Operation | Git ms / KiB | Grit before ms / KiB | Candidate ms / KiB |
| --- | ---: | ---: | ---: |
| walk-50000-no-graph | 90.06 / 19,020 | 181.79 / 53,248 | 173.38 / 53,612 |
| log-50000-no-graph | 111.72 / 19,348 | 416.33 / 102,740 | 415.80 / 102,904 |
| walk-50000-graph | 27.32 / 13,652 | 47.62 / 37,552 | 48.62 / 37,820 |
| log-50000-graph | 111.00 / 22,472 | 407.45 / 105,276 | 407.77 / 105,548 |
| log-100-no-graph | 2.25 / 6,664 | 3.35 / 12,268 | 3.32 / 12,520 |
| log-real-grit-main | 32.26 / 16,756 | 73.68 / 30,872 | 67.94 / 30,928 |

Packed-read Criterion timings below are means. Each ODB read inflates an object through the public API under a pack read context, bypassing the small-object result cache. The delta case invalidates the owning ODB's pack caches and checks an explicitly constructed 80-edge chain. The history fixture now distinguishes genuinely loose objects from Git-packed objects and asserts its storage layout. Graph use is enabled/disabled independently. The before/after generated history packs had identical checksums.

| Library operation | Before µs | Candidate µs | Change |
| --- | ---: | ---: | ---: |
| `packed_odb_read/blob` | 1.61 | 1.46 | -9.7% |
| `packed_odb_read/commit` | 1.42 | 1.22 | -14.5% |
| `packed_odb_read/delta` | 85.81 | 82.38 | -4.0% |
| `packed_odb_read/tree` | 0.48 | 0.37 | -22.8% |
| `packed_object_read/deep_delta_chain` | 38.58 | 35.45 | -8.1% |
| `packed_object_read/whole` | 1.34 | 1.25 | -6.6% |
| `revwalk_topo/loose/commit_graph` | 6147.32 | 5839.02 | -5.0% |
| `revwalk_topo/loose/no_commit_graph` | 68354.85 | 68161.40 | -0.3% |
| `revwalk_topo/packed/commit_graph` | 5856.66 | 5838.78 | -0.3% |
| `revwalk_topo/packed/no_commit_graph` | 32295.20 | 33906.03 | +5.0% |

The retained object suite benchmarks a reusable flate2 decoder on small commit/tree payloads, 4 KiB and 1 MiB blobs, and delta instructions; the reference candidate also measured Sirocco. The 1 MiB case is highly compressible; it is not a prediction for arbitrary blobs. The existing direct `ZlibDecoder` microbench remains available. The full `revwalk` suite also records author-filtered walks, bounded walks, counts, and merge-base, with time and RSS against system Git.

Historical raw data remains in reference candidate `8a0f13f4` under `grit-utils/baselines/`: `sirocco-process-comparison.json`, `sirocco-process-comparison-unpinned.json`, `sirocco-criterion.json`, and the before/candidate `sirocco-revwalk-*.json` reports. These experimental reports are excluded from the current baseline tables. Compression is separate follow-up work requiring its own wall-time evidence.

Reproduce with a frozen binary for each revision:

```bash
cargo build --profile dist -p grit-cli -p grit-utils
cargo bench -p grit-lib --bench objects -- 'zlib_reused|packed_odb_read|packed_object_read' --sample-size 20 --warm-up-time 1 --measurement-time 3
cargo bench -p grit-lib --bench history -- revwalk_topo --sample-size 20 --warm-up-time 1 --measurement-time 3
target/dist/grit-bench revwalk --grit /path/to/frozen/grit --warmup 2 --min-runs 5 --format json --output revwalk.json
```

## Undoing a commit without a commit-graph (2026-10-27)

On `git/git` (82,605 commits, 1,012 tags), `reset --soft HEAD~1` after an empty
local commit now stays near the commits being left behind. It reuses packed-ref
peel records and stops the advisory orphan paint by date when no commit-graph
exists. With a graph it keeps the exact generation stop; log, revision listing,
merge-base and transfer walks keep their existing exact behavior.

Factory VM, four vCPUs, Rust 1.99.0, `release` profile, system Git 2.55.0.
Both CLI binaries use `cargo build --release -p grit-cli`. One warmup and nine
interleaved runs per tool use Python `perf_counter` around `/usr/bin/time -v`;
preparation is excluded. Builds and tests are stopped during measurement, and
automatic Git maintenance is disabled in the fixture.

| Tool | Median wall (ms) | Peak RSS (MiB) |
| --- | ---: | ---: |
| System Git | 4.14 | 6.61 |
| Main (`3c1a41ef2`) | 703.00 | 86.87 |
| Bounded orphan report | 23.18 | 46.14 |

The committed `daily-reset-undo-commit-before.json` and
`daily-reset-undo-commit-after.json` compare reset and the existing log and
revision-count scenarios on the canonical `v2.47.0` fixture (75,031 commits):

| Scenario | Main (ms) | Branch (ms) | Git before / after (ms) |
| --- | ---: | ---: | ---: |
| Undo one commit | 722.30 | 20.15 | 3.64 / 3.62 |
| Log, 100 JSON commits | 6.82 | 5.84 | 4.08 / 3.90 |
| Revision count | 813.38 | 810.33 | 574.15 / 558.79 |

Six additional paired, interleaved `/usr/bin/time -v` samples per CLI check the
canaries' process memory and timing:

| Operation | Main (ms / MiB median RSS) | Branch (ms / MiB median RSS) |
| --- | ---: | ---: |
| Log, 100 JSON commits | 6.79 / 20.92 | 6.25 / 21.66 |
| Revision count | 789.12 / 124.68 | 785.97 / 125.00 |

The log canary's median RSS is about 0.75 MiB higher in this sample; full revision
counting differs by about 0.32 MiB. These walks retain their exact algorithms
and gain no per-commit data from the orphan-report stop mode.

Without a graph, the orphan report follows a date floor and a five-commit slop
window. Severe clock skew can make its count inaccurate; it is advisory and
never controls repository writes. See [reset](https://grit-scm.com/docs/benchmarks/commands/reset.md/index.md) and
[switch](https://grit-scm.com/docs/benchmarks/commands/switch.md/index.md).

## Undoing a commit with many tags (2026-10-28)

`reset/undo-commit@git-git` was still about 6× slower than Git, and used about 5.4× its
memory. The orphaned-commit report paints from every ref, and on the `v2.47.0` fixture
(954 tags, no commit-graph) it had to learn the date of each tip commit before taking a
step. Each tip cost a commit-graph existence probe, a header probe, and then a full read
and inflate through the pack mapping. Kernel fault-around on those scattered reads left
about 20 MB of the pack resident. The ancestry pre-check also walked a hundred commits
below the undone commit's date.

Now a paint with 64 or more tips on one side reads each packed tip commit's header block
with one positioned read, on scoped worker threads, and records its parents and date
before painting. Tags, loose or deltified commits and unusually long headers keep the full
read. The pre-check first looks for the target on the undone commit's first-parent chain,
which settles `HEAD~n` without the date-slop walk. The counts are unchanged.

Factory VM, four vCPUs, Rust 1.99.0, `release` profile, system Git 2.43.0. Main is
`e91d222f6`. Median of 20 runs per tool; the target was measured in two interleaved
rounds:

| Tool | Round 1 (ms / MiB) | Round 2 (ms / MiB) |
| --- | ---: | ---: |
| System Git | 2.13 / 8.2 | 2.01 / 8.1 |
| Main | 12.06 / 46.3 | 11.96 / 46.2 |
| Branch | 6.58 / 29.2 | 6.93 / 29.2 |

The committed `daily-reset-undo-commit-5128-before.json` (main) and
`daily-reset-undo-commit-5128-after.json` (branch) cover every `reset/*` and
`switch/*` scenario. They also cover the fetch, push, log, branch, merge-base, ahead and
range-count scenarios that paint against many refs (10 runs each). Apart from the
target, every scenario stays within run-to-run noise. Repeated interleaved runs of the
largest apparent moves (`branch/merged`, `log/json-n10-loose-tags`, `switch/branch` and
`reset/hard`) flipped sign between rounds.

Without tags the same reset takes about 3 ms. Of the remaining gap to Git, roughly 0.45 ms
lists the 954 refs, the parallel header reads take 1.6 to 1.9 ms (mostly per-stream inflate
setup), and process start-up and the ref and reflog update take the rest.

## Undoing a commit after the tags were packed (2026-10-31)

Two costs were left in `reset --soft HEAD~1` on `git/git` without a commit-graph. The
orphaned-commit report still dated all 947 packed tag tips before its first step, which read
every tagged commit's header and faulted most of the 10 MB pack index into memory. And reset
logged each reflog line in a ref transaction of its own, which loads the whole reflog. That made
every reset cost grow with `HEAD`'s history, about 16 ms of reflog work at 24,000 lines.

A ref recorded in `packed-refs` names a commit that existed when the file was written, so
within the one-day clock-skew allowance the report already relies on, it can only reach commits
dated at most a day after that. The report now hides only the target, loose refs, the stash and
other worktrees' `HEAD`s first. It keeps that answer when every commit left behind is dated more
than a day after `packed-refs` was last written. Otherwise it stops at the first older commit and
runs the full walk, so the counts match the full walk. Reset now moves the ref in one
compare-and-swap transaction and appends the reflog lines, as merge does.

Factory VM, four vCPUs, Rust 1.99.0, `release` profile, system Git 2.43.0. Main is
`3c14b849f`. Hyperfine, 5 warmups and 30 runs per tool with the commit to undo made in
`--prepare`. Peak RSS is the median of five `/usr/bin/time` runs. The fixture is a
`v2.47.0` clone with 947 packed annotated tags:

| Shape | Git (ms / MiB) | Main (ms / MiB) | Branch (ms / MiB) |
| --- | ---: | ---: | ---: |
| Fresh clone, 1-line `HEAD` reflog | 1.95 / 6.2 | 6.81 / 26.9 | 6.48 / 27.8 |
| Fresh clone, 2,000-line `HEAD` reflog | 1.94 / 6.2 | 9.41 / 28.1 | 7.53 / 27.5 |
| `packed-refs` 3 days old, 1-line reflog | 1.93 / 6.1 | 8.00 / 26.9 | 4.20 / 18.6 |
| `packed-refs` 3 days old, 2,000-line reflog | 1.81 / 6.1 | 8.89 / 27.7 | 4.04 / 18.3 |

The daily harness clones its fixture minutes before timing, so its `reset/undo-commit@git-git`
row is the first shape: the tags were packed within the skew allowance of the undone commit, and
the report still has to read them. The committed `daily-reset-undo-commit-5845-before.json`
(main) and `daily-reset-undo-commit-5845-after.json` (branch) cover every `reset/*` scenario
except `reset/blobless@git-git`, plus the `switch`, `log` and `tag` rows that list or peel
packed refs. `reset/blobless@git-git` fails its pre-timing check on this VM with main too. Apart
from the target (7.59 ms on main, 7.21 ms on the branch), every row stays within run-to-run noise.
The apparent move in `reset/sparse@sparse-unique-120k` reversed in an interleaved rerun (38.6 and
42.5 ms on main, 38.4 and 37.5 ms on the branch).

## Reading the tagged commits' headers in a batch (2026-11-02)

In the daily `reset/undo-commit@git-git` shape the tags are packed seconds before the undone
commit, so the orphaned-commit report still dates all 946 tagged commits before its first step.
Profiling showed where those header reads went. Each one looked the pack and its cached bytes up
by path (two SipHash lookups of the full path) and zeroed a 4 KiB read window. Its inflate buffer
was sized exactly to the object, and zlib only runs its fast decode loop with 258 bytes of output
room to spare, so most of every commit was decoded one symbol at a time. The scattered pack index
lookups also faulted in 8.6 MB of the 10 MB `.idx`, through kernel fault-around.

The tips are now read through a batch call that resolves the pack, its bytes and its object-count
check once per worker chunk and reuses its buffers. Inflate gets its spare output room. The first
read fetches 1 KiB of the entry, enough for 99.9% of these commits, and a second read takes the
full window only when that falls short. In an index of 4,096 objects or more, a batch lookup
reads the 128 object ids around the predicted row with a positioned read, after checking that the
file at the index path still has the mapped index's trailer. The matching offset comes from the
mapping, so only the offset table, a fifth the size of the object ids, can become resident. The
calling thread works on the first chunk while the other workers start, and the dates and parents
read are recorded under one lock. The report's counts are unchanged.

Factory VM, four vCPUs, Rust 1.99.0, `release` profile, system Git 2.43.0. Main is `285dbe512`.
`grit-bench daily`, 20 runs per tool, in four alternating rounds:

| Tool | Round 1 (ms / MiB) | Round 2 (ms / MiB) | Round 3 (ms / MiB) | Round 4 (ms / MiB) |
| --- | ---: | ---: | ---: | ---: |
| System Git | 1.99 / 8.1 | 2.03 / 8.3 | 2.00 / 8.3 | 1.96 / 8.2 |
| Main | 6.47 / 24.3 | | | 6.45 / 24.3 |
| Branch | | 6.09 / 19.0 | 6.15 / 19.3 | |

Of the remaining gap to Git, about 0.7 ms is process start-up (`grit --version` takes 1.4 ms,
`git --version` 0.7 ms), about 0.45 ms lists the refs, and the parallel header reads take
1.5 to 2 ms. Most of that is zlib building each commit's Huffman tables. Without a commit-graph,
reading the tagged commits cannot be avoided: Git's own orphan warning in `checkout` walks
`--not --all` the same way.

The committed `daily-reset-undo-commit-6751-before.json` (main) and
`daily-reset-undo-commit-6751-after.json` (branch) cover every `reset/*` scenario except
`reset/blobless@git-git`, plus the rows that paint against many refs or map patch-ids in parallel:
`branch/contains`, `branch/merged`, `tag/contains`, `fetch/noop`, `push/incremental`, `log`,
`merge-base`, `shortlog/ahead`, `switch/branch`, `pick/one` and `rebase/onto` (10 runs each).
Apart from the target, every row stays within run-to-run noise. Seven rows read 2 to 6% slower
in that first pass, while Git itself ran 1 to 3% slower in the branch's round: `branch/contains`,
`fetch/noop`, `push/incremental`, `tag/contains`, `merge-base/branches`, `rebase/onto` and
`log/json-n10-loose-tags`. In four alternating reruns `fetch/noop` was faster on the branch, and
the others were within 3.5% either way. Alternating single runs, 80 to 100 per binary, put
`merge-base v2.46.0 v2.47.0` at 3.78 ms on main and 3.80 ms on the branch, and `log -n 10 --json`
with 955 loose tags at 12.62 and 12.71 ms.

## Tree metadata diff (2026-10-27)

Issue [#2083](https://github.com/gitbutlerapp/grit/issues/2083) exposed a second,
unbounded parsed-tree cache in the tree diff walker. It pinned completed sibling
trees even after the ODB’s 65,536-entry cache evicted them. The walker now uses
that existing budget and retains only active recursion frames. Name-only and
name-status file records consume their input paths and modes instead of cloning
them. The public `DiffEntry` representation and owned result vector are unchanged.

The exact reproduction has 120,000 files across 400 directories, each initially
containing 20 lines; the second commit appends one line to 20,000 files. The
public API consumer opens the repository, resolves both commits, calls
`Diff::tree_to_tree`, and prints each status/path directly. CLI rows use
`grit diff --name-status base HEAD`, with default rename detection and with
`--no-renames`. Both are compared with system Git; every result contains the same
20,000 modified paths. A second copy stores all objects in a Git-written pack
(`git repack -ad` followed by `git prune-packed`). Both repositories pass
`git fsck --strict`.

Measurements use Rust 1.99.0 `release` builds on a four-vCPU AMD EPYC 7763 Linux VM
with 16 GiB RAM and stock Git 2.55.0. Frozen main is `b4e048f1c`.
After one warmup per tool, five interleaved runs alternate Git, frozen main, and
the fix, reversing order on alternate rounds. Builds and tests are idle.
Wall time is the median from Python `perf_counter` around `/usr/bin/time -v`;
RSS is the maximum process peak across the five runs, in KiB.

| Storage / operation | Git ms | Main ms | Fixed ms | Git RSS KiB | Main RSS KiB | Fixed RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| loose / public tree_to_tree consumer | 37.79 | 67.05 | 58.45 | 11,496 | 41,040 | 21,308 |
| loose / CLI default renames | 38.68 | 54.76 | 44.94 | 11,496 | 53,656 | 32,964 |
| loose / CLI --no-renames | 38.26 | 53.14 | 43.62 | 11,500 | 51,764 | 33,044 |
| packed / public tree_to_tree consumer | 25.74 | 60.02 | 51.01 | 22,760 | 50,852 | 30,224 |
| packed / CLI default renames | 26.51 | 50.32 | 39.41 | 22,752 | 64,668 | 43,888 |
| packed / CLI --no-renames | 25.98 | 48.09 | 37.71 | 22,744 | 63,132 | 41,864 |

The loose public API uses **48% less peak RSS** and the loose default CLI uses
**39% less**. Packed reductions are **41%** and **32%**, respectively. All six
end-to-end cases are **13–22% faster than main**. Git still uses less memory.
Raw samples, GNU time output, and Criterion estimates are committed in
[`tree-metadata-2083.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/tree-metadata-2083.json).

The isolated public-API allocation regression holds changed paths at 400 and
increases unchanged tree metadata from 51,200 to 204,800 files. Main’s peak live
heap grows from 11,528,650 to 44,477,899 bytes; the fix grows from 8,084,206 to
12,246,185 bytes. Result storage remains 111,712 bytes in both fixed cases. The
remaining peak includes the ODB’s bounded caches, larger individual loose tree
objects, and active directory frames. The owned result vector also scales with
changed paths, so these measurements do not imply a fixed RSS ceiling.

Criterion’s `diff/tree_diff_siblings` covers the same 120k/20k shape with distinct
sibling OIDs and with repeated child OIDs, using both cold and warm ODB handles.
With 1 s warmup, 2 s measurement, and 20 samples per case, distributed cold/warm
estimates change from 20.620/24.655 ms to 19.828/23.989 ms. Repeated cold/warm
estimates are 3.392/3.991 ms before and 3.493/3.932 ms after; Criterion detects no
significant change in either repeated-subtree case.

## Tree metadata diff through the CLI (2026-10-27)

After the library fix above, `grit diff --name-status base HEAD` on the same
fixture still peaked about 14 MB above the public `Diff::tree_to_tree` consumer.
Three copies were responsible. Rename detection, which the CLI runs by default,
split every entry into new vectors and rebuilt the list even when nothing could
pair, so a diff of modifications briefly held its metadata twice. The CLI then
converted the entries into library `FileDiff` rows and those into its own output
rows, keeping a third full list alive. Finally, each tree read for the ODB’s
parsed-tree cache was also copied into the raw loose-object cache.

Rename detection now returns its input unchanged, without reading content, when the
list has no added/deleted pair and its paths are already in order. The CLI maps each
entry straight to its output row through `FileDiff::from_metadata`, and trees read for
the parsed cache skip the raw copy. Diff lists are now ordered by raw path bytes in
one shared helper, `sort_entries_by_path`, which also skips the stable sort’s scratch
allocation when the list is already ordered. That moves non-UTF-8 paths to the
position Git gives them: previously, an encoded name sorted by its encoded form.

The setup matches the section above: the exact 120k/20k fixture, loose and packed
copies, Rust 1.99.0, a four-vCPU AMD EPYC 7763 VM with 16 GiB RAM, and stock Git
2.55.0. Main is `3681c1778`. Each tool gets one warmup, then five interleaved runs
that reverse order each round, with builds and tests idle. Wall time is the median;
RSS is the maximum process peak, in KiB. The Git column is `git diff --name-status`
with the same rename setting (`--no-renames` for the API row).
Release rows use `cargo build --release`. Dist rows use the `dist` profile that
releases ship (fat LTO, one codegen unit).

| Storage / operation | Git ms | Main ms | Fixed ms | Git RSS KiB | Main RSS KiB | Fixed RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| loose / public tree_to_tree consumer | 50.49 | 59.94 | 57.61 | 10,856 | 19,896 | 18,244 |
| loose / CLI `--name-status` | 50.48 | 55.31 | 52.87 | 10,952 | 33,624 | 30,480 |
| loose / CLI `--name-status --no-renames` | 50.64 | 55.98 | 53.91 | 10,888 | 33,652 | 30,604 |
| loose / CLI `--name-only` | 50.88 | 54.42 | 52.71 | 10,944 | 33,560 | 30,592 |
| loose / CLI `--name-status --json` | — | 61.25 | 56.93 | — | 33,640 | 30,480 |
| loose / CLI `--name-status`, dist | 50.48 | 52.06 | 50.37 | 10,952 | 30,884 | 27,480 |
| packed / public tree_to_tree consumer | 34.44 | 50.80 | 50.50 | 22,364 | 30,400 | 28,688 |
| packed / CLI `--name-status` | 34.74 | 50.66 | 46.99 | 22,236 | 44,348 | 41,380 |
| packed / CLI `--name-status`, dist | 34.74 | 47.57 | 45.15 | 22,236 | 41,460 | 38,124 |

Peak RSS falls by 3.0–3.4 MB for every CLI row (9–11%) and by 1.7 MB for the API
consumer, and no row gets slower. The live heap now matches the library’s: at exit, the CLI holds
10.4 MB of anonymous memory and the consumer 9.8 MB. The remaining gap to the
consumer, and most of the gap to Git, is executable code. The CLI runs about 0.9 MB
of distinct machine code for this command, but that code is spread across 162
separate 64 KiB blocks of the 24 MB text segment. The kernel maps whole fault-around
windows, so about 10 MB of the `grit` binary stays resident. Git touches 1.9 MB of
its binary, and the API consumer 2.4 MB. Fat LTO in the dist profile cuts this to
6.9 MB; thin LTO does not help. This cost is per process rather than per diff, so it
is a build-layout problem, not a diff-memory one: an empty `grit diff HEAD HEAD`
peaks at 15 MB.

Criterion’s `diff/tree_diff_siblings` gains a `distributed/cold+renames` case, the
CLI’s path with default rename detection. With 1 s warmup, 3 s measurement and 20
samples, it moves from 28.2–29.6 ms on main (three runs) to 26.3–27.2 ms (two runs).
The other four cases stay within run-to-run noise: distributed cold/warm 23.7–24.9/26.8–28.5 ms before
and 23.6–24.1/26.2–28.2 ms after, repeated cold/warm 4.28–4.57/4.59–4.71 ms before
and 4.53–4.68/4.58–4.85 ms after. Raw samples are in
[`tree-metadata-cli-2083.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/tree-metadata-cli-2083.json).

## Sparse reset and switch (2026-10-27)

Issue [#2081](https://github.com/gitbutlerapp/grit/issues/2081) found full-tree
expansion behind a compact sparse index. Reset and switch now retain matching
excluded tree entries when those directories are absent from disk. Target indexes
reuse those entries and their tree IDs, and hard reset carries its index through
checkout without an intermediate write and reload. Changed directories, present
out-of-cone paths and unmerged guard checks still use expanded rows for correctness.

Measured with Rust 1.99.0 `release` builds on a four-vCPU AMD EPYC 7763 Linux VM
with 16 GiB RAM and stock Git 2.55.0. Frozen main is `02e8900b78ea`. The sparse
fixtures have 300 visible files and 400 excluded directories (700 index rows),
with 300 visible changes between two commits. The unique-blob variant additionally
changes 5,000 excluded paths. The populated control tracks 10,000 files.

Each value is the median of ten endpoint samples: five rotated rounds per tool,
with a base/edited flip in each round after one warmup. Python `perf_counter`
wraps `/usr/bin/time -v`; RSS is in KiB. Builds and tests were paused. Both grit
binaries were copied once to fresh files using the same method before sampling.

| Fixture / operation | Git ms | Main ms | Fixed ms | Git RSS KiB | Main RSS KiB | Fixed RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| sparse120k reset | 30.18 | 1063.71 | 30.66 | 10,514 | 73,548 | 17,592 |
| sparse120k switch | 38.72 | 461.37 | 31.86 | 6,096 | 37,388 | 18,040 |
| sparse240k reset | 36.46 | 2251.56 | 30.65 | 15,420 | 141,134 | 17,602 |
| sparse240k switch | 52.36 | 1021.34 | 32.15 | 6,396 | 56,614 | 18,004 |
| unique120k_changed5000 reset | 41.08 | 1152.15 | 534.75 | 18,320 | 110,460 | 108,318 |
| unique120k_changed5000 switch | 52.06 | 556.55 | 446.04 | 13,838 | 76,252 | 75,890 |
| full10k reset | 45.74 | 57.67 | 53.58 | 7,976 | 22,740 | 22,570 |
| full10k switch | 38.92 | 32.11 | 32.11 | 8,016 | 19,104 | 19,340 |

At 240k paths, reset improves **73.5×** and switch **31.8×**. Doubling unchanged
excluded paths leaves grit near 30–32 ms and 17–18 MiB. The populated control
keeps its timing and memory profile: reset is faster, while switch differs by
less than 0.01 ms and less than 0.3 MiB in this run. All repositories pass system Git's
status, write-tree and strict fsck checks after the grit operations.

The unique-blob case documents the remaining fallback cost: changed excluded
directories still expand. Reset improves about **2.2×**, and switch about **1.2×**,
with peak memory essentially unchanged. Hard reset now keeps changed excluded trees
opaque too (see [Hard reset of changed excluded trees](#hard-reset-of-changed-excluded-trees-2026-10-28));
switch still expands them.

The fixtures and library benchmarks live in `grit-test-support/src/sparse_scale.rs`
and `grit-lib/benches/hot_paths.rs`. Daily rows are
`reset/sparse@sparse-120k`, `reset/sparse@sparse-240k`,
`switch/sparse@sparse-120k`, `switch/sparse@sparse-240k`, and the corresponding
`sparse-unique-120k` variants. Each daily iteration measures both endpoints;
fixture setup and correctness checks are outside the timed interval.

Raw commands, GNU time output, samples and verification output:
[`sparse-transitions-2081.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/sparse-transitions-2081.json).

### Parent replacement review follow-up

Sparse cleanup now retains only paths still tracked in the target and checks for
symlinked parents. Deleted descendants stay in checkout's deletion phase, before
a parent can be replaced by a file or symlink. Git-control regressions check both
replacements, reset-to-HEAD with a symlinked parent, an external sentinel, and Git
operations after grit writes. Sparse round trips compare file contents as well as
paths and index state.

A fresh comparison on the follow-up VM used main `e98de8a16348`, pre-review
`283078a7f`, and the fix, with stock Git 2.55.0 and Rust 1.99.0 release builds.
Values are median **milliseconds / peak KiB** from ten endpoint samples per
operation, after a warmup pair, with tool order rotated over five rounds. Builds
and tests were paused. The full-checkout control has 10,000 visible files and 300
edits. All tools retain the expected worktree/index shape and pass Git status and
strict fsck. These are a separate run from the initial measurements above.

| Fixture / operation | Git | Current main | Pre-review | Fixed |
| --- | ---: | ---: | ---: | ---: |
| sparse120k reset | 22.87 / 10,642 | 765.78 / 75,320 | 35.01 / 17,884 | 35.00 / 17,684 |
| sparse120k switch | 29.94 / 6,224 | 333.93 / 37,122 | 35.80 / 18,150 | 35.55 / 17,974 |
| sparse240k reset | 28.03 / 15,480 | 1588.11 / 141,744 | 34.92 / 17,720 | 34.85 / 17,880 |
| sparse240k switch | 40.39 / 6,524 | 695.74 / 57,146 | 35.47 / 18,198 | 35.71 / 18,024 |
| unique120k_changed5000 reset | 32.27 / 18,406 | 829.94 / 110,646 | 404.66 / 108,354 | 403.64 / 108,456 |
| unique120k_changed5000 switch | 40.96 / 13,940 | 412.18 / 76,180 | 349.12 / 76,364 | 339.30 / 76,552 |
| full10k reset | 35.09 / 7,894 | 55.74 / 23,476 | 52.32 / 23,178 | 52.38 / 22,922 |
| full10k switch | 30.82 / 8,146 | 36.06 / 19,024 | 36.14 / 19,562 | 37.11 / 19,348 |

The cleanup fix preserves the pre-review timing and memory profile, including
changed excluded trees and full checkouts. Endpoint samples and binary hashes are
in [`sparse-transitions-review-2081.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/sparse-transitions-review-2081.json).

### Hard reset of changed excluded trees (2026-10-28)

`reset/sparse@sparse-unique-120k` changes 5,000 excluded paths, so all 400 excluded
directories differ between the two commits. Hard reset used to expand each of them
into full index rows (about 120,000), diff and probe them, and collapse them again
when writing the index. A hard reset discards staged content and never writes
excluded paths, so an excluded directory that is absent from disk now keeps its
sparse-directory entry whenever the target still has a tree there, and the entry
takes the target's subtree ID, as Git's sparse-index reset does. The tree diff
stops at those directories instead of reading both sides' subtrees. Directories
present on disk, and directories the target removes or replaces with a file, are
still expanded. Switch keeps the expanded path because it must check for staged
changes.

Daily-matrix medians of ten runs (each run resets out and back), measured in one
sitting on a four-vCPU Intel Xeon Linux VM with stock Git 2.43.0 and Rust 1.99.0
release builds. Values are **milliseconds / peak MB**:

| Scenario | Git | Before | After |
| --- | ---: | ---: | ---: |
| `reset/sparse@sparse-unique-120k` | 56.4 / 23.0 | 928.6 / 94.9 | 30.9 / 22.0 |
| `reset/sparse@sparse-120k` | 28.4 / 12.7 | 27.1 / 20.9 | 24.4 / 21.1 |
| `reset/sparse@sparse-240k` | 40.2 / 17.7 | 25.1 / 20.8 | 31.3 / 21.4 |
| `reset/hard@synth-10k` | 15.5 / 11.2 | 25.5 / 25.6 | 24.9 / 26.0 |
| `reset/hard-delete-heavy@wide-48k` | 226.2 / 22.2 | 256.2 / 53.3 | 246.4 / 53.8 |

The changed-tree reset is **30×** faster and uses a quarter of the memory. The other
rows run no new code apart from an index scan for sparse entries, and stay within noise. An
interleaved `hyperfine` rerun of the 240k round trip (40 runs) measured 28.9 and
33.7 ms before against 27.6 and 29.5 ms after (Git: 44.3 and 42.8 ms). Full reports
for every reset scenario and the sparse switch rows:
[`daily-reset-sparse-unique-5074-before.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-reset-sparse-unique-5074-before.json)
and
[`daily-reset-sparse-unique-5074-after.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-reset-sparse-unique-5074-after.json).

## Sparse stash apply/pop (2026-10-27)

Issue [#2082](https://github.com/gitbutlerapp/grit/issues/2082) reproduced with
120,000 unique tracked files across 400 directories, 20 lines per file, and 3,000
appended edits saved by system Git. Stash apply/pop now shares one changed-path
plan between overwrite checks and application, skips identical subtree OIDs,
and mutates one loaded index. Moved-HEAD metadata is looked up only for affected
paths through the bounded tree cache. Sparse checkout replacements reuse existing
index slots, and stash scratch data is released before index serialization.

Measured on a four-vCPU AMD EPYC 7763 Linux VM with 16 GiB RAM, Rust 1.99.0 and
stock Git 2.55.0. Frozen main is `02e8900b78ea`. The original copy-and-edit
reproduction first confirmed 85,368–86,132 KiB on main versus Git's
37,468–37,752 KiB on warm runs. The fix uses 49,068–49,348 KiB and takes
223–246 ms versus main's 490–525 ms and Git's 640–694 ms in those warm runs.
Worktree diffs match byte for byte and strict fsck succeeds for both tools.

The scaling comparison below keeps the same 3,000 edits while selecting 100,
200, or 400 directories from that fixture. The smaller fixtures reuse the
original loose objects through Git alternates; compare tools within each row,
since alternate lookups add a fixed per-edit cost to the smaller fixtures.
Git, frozen main, and the fix run in alternating order on the same checkout. Each tool/output gets one warmup and
three recorded runs with builds and tests paused. Times are medians of Python
`perf_counter` around `/usr/bin/time -v`; RSS is the maximum recorded peak, in
KiB. Git controls use ordinary `git stash pop` for every output format. Worktree
and staged diffs, index stages, status, empty stash state, and strict fsck are
checked against Git. These are `release` builds:

| Tree / output | Git ms | Main ms | Fixed ms | Git RSS KiB | Main RSS KiB | Fixed RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| 30k / human | 401.95 | 336.29 | 293.72 | 13,460 | 37,128 | 27,760 |
| 60k / human | 473.44 | 391.12 | 277.92 | 21,612 | 53,816 | 36,112 |
| 120k / human | 606.98 | 439.17 | 245.69 | 37,772 | 89,308 | 49,312 |
| 120k / json | 600.74 | 481.24 | 226.77 | 37,772 | 89,340 | 49,320 |
| 120k / markdown | 593.41 | 448.21 | 213.78 | 37,900 | 89,432 | 49,436 |

The same three-way comparison with shipped `dist` builds:

| Tree / output | Git ms | Main ms | Fixed ms | Git RSS KiB | Main RSS KiB | Fixed RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| 30k / human | 401.29 | 320.43 | 267.41 | 13,520 | 33,784 | 23,896 |
| 60k / human | 475.79 | 365.38 | 284.92 | 21,796 | 50,380 | 30,296 |
| 120k / human | 615.53 | 444.33 | 217.25 | 37,708 | 85,356 | 45,152 |
| 120k / json | 614.14 | 454.24 | 224.51 | 37,584 | 85,324 | 45,004 |
| 120k / markdown | 603.32 | 459.39 | 229.27 | 37,772 | 85,292 | 44,872 |

A sparse-switch control over the same 120k-file tree (3k changed paths) also
improves median time: release **175.70 → 170.30 ms**, dist **187.19 → 170.21 ms**,
versus Git **409.22 / 412.13 ms**. Maximum RSS is **48,152 → 50,004 KiB**
(release) and **44,404 → 45,320 KiB** (dist), small process/allocator-residency
differences that remain in the raw samples. No worktree/index mismatch occurs.

Raw samples, GNU time output, binary hashes and reproduction scripts:
[`stash-pop-2082.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/stash-pop-2082.json).

Heaptrack on the original loose-object fixture measures peak live heap falling
from **69.53 to 29.28 MB**, with allocation calls falling from **2,558,667 to
373,732**. At the old peak, flattened-tree allocations account for 11.79 MB,
the separately retained parsed index for 15.36 MB, and checkout's entry-vector
growth for 26.88 MB. The serializer still needs its output buffer; one index and
process/allocator overhead remain, so total RSS remains above Git.

The Linux Rust RSS regression exercises human, JSON, and Markdown pop in fresh
processes at 4k and 120k files with the same 3k edits. It bounds tree-size memory
growth against Git to control for debug/process overhead, retains an absolute
Git-relative bound in optimized builds, and compares state and bytes. Library
compatibility tests cover staged additions/deletions, moved HEAD, conflicts,
untouched local changes, filters, symlinks and overwrite protection. Criterion
adds `stash/pop_sparse_wide/100k_files_3k_edits`; the daily matrix adds
`stash/pop-sparse@wide-120k` (120k files/3k edits; 4k/100 in smoke mode).

```bash
cargo test -p grit-cli --test stash_pop_memory -- --nocapture
cargo test -p grit-lib --test stash_git_compat --test stash_apply_git_compat \
  --test stash_apply_outcome_git_compat --test stash_apply_autocrlf
cargo bench -p grit-lib --bench hot_paths -- stash/pop_sparse_wide --test
```


### Raw filename identity recheck

The sparse plan now keeps raw Git path bytes through validation, overwrite
preflight, tree lookup, filesystem access and index reconciliation. This prevents
literal `grit+pathhex:` filenames from being mistaken for encoded diff paths.
The shared index-merge and rename helpers also decode diff paths before lookup.
Eight Git round-trip regressions cover overwrite rejection with stash retention,
clean apply/pop, deletion, moved HEAD, conflicts, attributes, staged changes and
staged renames, including non-UTF-8 filenames on Unix.

A fresh `release` comparison on the same VM uses current main `ea7901dc43cc`,
the reviewed optimization `e16af2a43`, and the path correction. The fixture follows
the original reproduction (120k unique loose blobs, 3k appended edits). Builds
and tests are paused during measurement. Each tool/output receives one warmup
and three recorded runs; wall time is the median and RSS is the largest warm
peak. Every run compares diffs, index stages, status and stash state with Git;
strict fsck passes. Git uses its normal human output in each control.

| Output | Git ms | Main ms | Reviewed ms | Revised ms | Git RSS KiB | Main RSS KiB | Reviewed RSS KiB | Revised RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| Human | 606.54 | 488.70 | 249.70 | 242.31 | 37,736 | 84,500 | 46,316 | 46,088 |
| JSON | 606.19 | 468.34 | 251.11 | 240.33 | 37,712 | 84,268 | 46,004 | 45,940 |
| Markdown | 594.70 | 476.74 | 235.87 | 248.84 | 37,740 | 84,320 | 46,164 | 46,300 |

The correction retains the memory reduction and fast warm runtime. One human
run took 1.81 seconds; it remains in the raw samples and median calculation.
Raw GNU time records, binary/source hashes, scripts and the exact overwrite
reproduction in human/JSON/Markdown are in
[`stash-pop-2082-path-revision.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/stash-pop-2082-path-revision.json).

## Sparse index collapse on write (2026-10-28)

`switch/sparse@sparse-unique-120k` changes 5,000 excluded paths, so switching
expands all 400 excluded directories (about 120,000 index rows) and folds them
back into sparse-directory entries when it writes the index. That fold removed
each directory's rows with a pass over the whole index and inserted its entry
into the sorted list, so its cost grew with directories × rows. It also re-read
the root tree for every directory. Writing a sparse index now decides every
directory first, against the index as it looks with the directories collapsed so
far, and rebuilds the entry list once. Which directories collapse is unchanged.

Daily-matrix medians of ten runs (each run goes out and back), measured in one
sitting on a four-vCPU Intel Xeon Linux VM with stock Git 2.43.0 and Rust 1.99.0
release builds. Values are **milliseconds / peak MB**:

| Scenario | Git | Before | After |
| --- | ---: | ---: | ---: |
| `switch/sparse@sparse-unique-120k` | 76.0 / 17.6 | 754.6 / 56.0 | 295.3 / 56.2 |
| `reset/sparse@sparse-unique-120k` | 54.5 / 21.8 | 911.6 / 90.9 | 453.2 / 91.6 |
| `switch/sparse@sparse-120k` | 42.3 / 8.0 | 25.5 / 20.7 | 25.8 / 20.8 |
| `switch/sparse@sparse-240k` | 62.6 / 8.1 | 26.0 / 20.6 | 26.2 / 21.2 |
| `reset/sparse@sparse-120k` | 29.3 / 12.1 | 26.0 / 20.7 | 27.1 / 21.3 |
| `reset/sparse@sparse-240k` | 38.6 / 16.9 | 26.5 / 20.8 | 26.4 / 21.2 |
| `stash/pop-sparse@wide-120k` | 361.3 / 53.6 | 126.5 / 49.1 | 123.8 / 49.5 |
| `switch/branch@synth-100k` | 161.2 / 35.9 | 84.3 / 43.3 | 89.4 / 43.3 |
| `switch/branch@synth-10k` | 22.5 / 10.2 | 18.0 / 21.7 | 18.4 / 22.1 |
| `switch/delete-heavy@wide-48k` | 124.6 / 21.8 | 132.4 / 33.8 | 130.6 / 34.1 |
| `switch/large-blob@big-asset` | 849.6 / 306.7 | 782.5 / 19.3 | 784.6 / 20.2 |

The changed-tree switch is **2.6×** faster and hard reset on the same fixture
**2×**. The remaining gap to Git is the expansion and the tree diff before the
collapse, and those rows dominate peak memory too. The other rows never collapse
a changed directory and stay within noise: two interleaved reruns of
`switch/branch@synth-100k` measured 88.7 and 88.5 ms before against 91.9 and
88.4 ms after. Full reports:
[`daily-switch-sparse-unique-5096-before.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-switch-sparse-unique-5096-before.json)
and
[`daily-switch-sparse-unique-5096-after.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-switch-sparse-unique-5096-after.json).

## Symmetric revision counts (2026-10-27)

Resolving `HEAD~500` parsed its parent chain, then the symmetric-range walk read
those commits again for the same parents and timestamps. Symmetric range
expansion now seeds the repository's existing metadata cache, capped at 1,024
entries. Reuse is guarded when shallow boundaries, promisor state, grafts or
replacements can change parent semantics; standalone resolution keeps its original
retention policy.

Measured against system Git 2.55.0 on a four-vCPU AMD EPYC 9V45 Linux VM with
16 GiB RAM and Rust 1.99.0 `release` builds. Frozen main is `aa1964221e04`.
The production deep-history fixture has 50,000 commits and 20,000 files; git/git
is pinned to v2.47.0 (`777489f9e09c`). All builds and tests
were stopped during measurements. Both grit executables were copied using the
same buffered-write method and warmed equally.

The table uses **30 rotated, interleaved runs per tool** after three warmups.
Wall time is the median from Python `perf_counter_ns` around `/usr/bin/time -v`,
including the helper's launch. Memory is the median process peak RSS in KiB.
The target improves **16.1%**; peak RSS is essentially unchanged.
Heaptrack records **18,591 → 13,586 allocation calls (26.9% fewer)**, with peak
heap usage remaining about 0.64 MB. The scoped cache avoids extra retention in
the standalone resolution controls.

| Scenario | Git ms | Main ms | Fixed ms | Main peak KiB | Fixed peak KiB |
| --- | ---: | ---: | ---: | ---: | ---: |
| `revs/left-right-count@deep-history` | 3.054 | 6.235 | 5.229 | 15,208 | 15,172 |
| `revs/author-n50@deep-history` | 2.414 | 4.559 | 4.479 | 16,884 | 16,960 |
| `revs/count-author@deep-history` | 99.796 | 127.828 | 127.825 | 25,566 | 25,540 |
| `revs/author-n50@deep-history-cg` | 2.551 | 4.359 | 4.305 | 19,524 | 19,518 |
| `revs/count-author@deep-history-cg` | 106.614 | 129.287 | 128.597 | 25,722 | 25,710 |
| `revs/count@git-git` | 424.569 | 575.766 | 579.342 | 125,950 | 125,962 |
| `revs/count-path-head@git-git` | 320.630 | 343.816 | 344.814 | 270,582 | 270,540 |
| `resolve/ancestor-500@deep-history` | 2.767 | 4.660 | 4.676 | 14,824 | 14,838 |
| `resolve/ancestor-10000@deep-history` | 16.161 | 28.993 | 29.148 | 16,296 | 16,216 |
| `resolve/head@deep-history` | 2.000 | 3.559 | 3.587 | 12,676 | 12,564 |
| `revs/left-right-first-parent@deep-history` | 3.060 | 6.297 | 5.242 | 15,208 | 15,096 |
| `log/oneline-100@git-git` | 3.830 | 5.175 | 5.078 | 21,120 | 21,008 |
| `show/head@git-git` | 2.605 | 3.899 | 3.837 | 16,324 | 16,276 |

A targeted direct-spawn daily recheck measures the target at
**5.822 → 4.361 ms**
(20 runs per build; system Git **2.643 → 2.345 ms**).
The earlier full-matrix captures experienced host drift, so the interleaved table
is the primary before/after comparison.
All measured grit outputs match frozen main byte for byte, and every count
matches system Git. Library tests cover reuse, standalone retention, invalid
objects, shallow and replacement guards, plus Git/grit object round trips and
navigation through loose objects, packs, commit graphs, tags and merged histories.

The registered `revs/count-path@git-git` row currently references `7e34d81`, absent
from the pinned fixture, so system Git rejects it. The table labels the working
`HEAD` path-count control separately.

Raw samples, commands, binary/source hashes and profiling counts:
[`revs-navigation-4759-paired.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/revs-navigation-4759-paired.json).
The standard runner captures are
[`daily-revs-navigation-4759-before.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-revs-navigation-4759-before.json)
and
[`daily-revs-navigation-4759-after.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-revs-navigation-4759-after.json).

## Inexact rename detection (2026-10-27)

Issue [#1683](https://github.com/gitbutlerapp/grit/issues/1683) reproduced on
`git/git`: line hashes were cached per blob, but every inexact candidate pair
still allocated and populated a new hash map. Rename and copy detection now
sort and combine byte weights once per blob and intersect the immutable tables
without allocating per pair. Index advancement avoids unpredictable hash-order
branches. Line hashing, CRLF normalization, scores, thresholds and tie-breaking
are unchanged.

Measured on a four-vCPU AMD EPYC 7763 64-Core Processor Linux VM with 16 GiB RAM, Rust 1.99.0 `release`
builds and stock Git 2.55.0. Frozen `origin/main` is `80dc322f7c2f`.
The upstream fixture is `git/git` at `37b3475fb530` (82,605 commits), comparing
`v2.40.0` with `master`. The second fixture moves 1,000 documents containing
100 unique lines each and edits one line per document, with
`diff.renameLimit=2000`.

After one warmup per tool, five interleaved runs alternate system Git, frozen
main and the fix, reversing order on alternate rounds. Builds and tests are
paused. Wall time is the median measured with Python `perf_counter` around
`/usr/bin/time -v`; RSS is the maximum process peak across those runs, in KiB.
The name-status cases use `-M`; the other tag cases use the listed output mode
with default rename detection unless marked `no-renames`.

| Scenario | Git ms | Main ms | Fixed ms | Git RSS KiB | Main RSS KiB | Fixed RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| name-status | 313.5 | 1248.9 | 180.1 | 56,616 | 79,828 | 79,620 |
| name-only | 312.7 | 1594.8 | 208.4 | 56,560 | 80,320 | 80,508 |
| numstat | 898.9 | 2308.4 | 944.0 | 118,724 | 171,336 | 170,220 |
| stat | 900.2 | 2312.8 | 922.6 | 118,896 | 171,064 | 169,916 |
| patch | 1179.0 | 2730.9 | 1352.4 | 118,528 | 203,728 | 203,992 |
| name-only-no-renames | 8.8 | 13.8 | 14.2 | 26,852 | 38,164 | 38,292 |
| numstat-no-renames | 662.6 | 728.9 | 721.5 | 119,188 | 158,940 | 159,096 |
| inexact-1000 | 319.8 | 5549.2 | 642.8 | 13,996 | 47,024 | 46,836 |

The tag name-status comparison is **6.9× faster than main** and **0.57× Git's
time**. The 1,000-file case improves **8.6×**, to **2.01× Git's time**. Rename-only
peak RSS remains essentially unchanged. The candidate matrix still uses more
memory than Git on the 1,000-file fixture; this change removes per-comparison
allocations without changing candidate storage. No-renames times differ by at
most 0.4 ms for name-only and 1.1% for numstat in this run.

Every output mode is byte-for-byte identical between main and the fix. Both
grit builds report 937 renames for the tag pair versus Git's 936, preserving the
existing similarity heuristic. All 1,000 synthetic rename path pairs match Git.
Rust tests check weighted line multiplicities, unterminated lines, binary bytes,
symmetry, table reuse and CRLF handling. An allocation regression covers both
rename and copy matrices against Git-written objects and checks rename mappings
with system Git.

Raw samples, complete GNU time output, commands, fixture generation and output
hashes: [`rename-scoring-1683.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/rename-scoring-1683.json).
Criterion mean estimates improve from **90.79 to 9.42 ms** for 64 documents
with shared paragraphs and **394.72 to 46.67 ms** for 256 documents with unique
paragraphs. Both binaries run with builds and tests paused (3-second warmup,
10 samples). The existing 2,000 × 2,000 matrix remains in the suite:

```bash
cargo bench -p grit-lib --features http-ureq,tracing --bench diff_renames -- diff_renames_adoc_similar
```

## Scattered pack-index lookups and log memory (2026-10-27)

Large pack indexes now predict an object's row from its hash prefix, then bracket
it with exponentially growing steps and finish with binary search. This avoids
reading both ends of a fanout bucket and dividing wide integers on each probe.
Every result still compares the complete object id; clustered hashes, missing ids,
SHA-1/SHA-256 and v1/v2 indexes retain their existing behavior. Bracketing and
binary search bound the probe count logarithmically.

On a 4-vCPU AMD EPYC 7763 Linux VM with 16 GiB RAM, Rust 1.99.0 `release` builds
and system Git 2.55.0, the pinned git/git fixture's 100-commit log uses **1,798 KiB
less peak RSS (8.2%)**. Headline elapsed time is essentially unchanged: a separate
300-run Hyperfine check measured 5.132 → 5.126 ms (Git: 3.606 ms). Startup and object decompression remain larger costs.

The table uses 20 rotated, interleaved runs per binary after three warmups, with
builds and tests stopped. Frozen main is `7cc7979535ce`; fixed output matches it
byte for byte in every row. Python `perf_counter` wraps `/usr/bin/time -v`, so these
wall times include that helper's launch; RSS is median peak RSS in KiB. Executables
were copied with the same buffered-write method and warmed equally. Mapped code
and file pages introduce RSS variation; the small cat/status differences are below
0.3 MiB. The path-limited row's roughly 1% timing difference did not persist in a
separate 100-run check (22.3 → 21.9 ms).

| Scenario | Git ms | Main ms | Fixed ms | Main KiB | Fixed KiB |
| --- | ---: | ---: | ---: | ---: | ---: |
| `log/all@git-git` | 53.14 | 73.62 | 73.56 | 35,422 | 33,894 |
| `log/author-n50@deep-history` | 3.13 | 5.49 | 5.43 | 17,124 | 15,888 |
| `log/author-n50@deep-history-cg` | 3.51 | 5.65 | 5.60 | 19,230 | 18,122 |
| `log/author@git-git` | 3.98 | 6.65 | 6.50 | 21,232 | 19,000 |
| `log/json-100@git-git` | 4.73 | 7.01 | 6.88 | 22,752 | 20,582 |
| `log/json-filter@git-git` | 3.89 | 7.20 | 7.19 | 21,620 | 19,630 |
| `log/limit-path@git-git` | 25.51 | 22.02 | 22.23 | 36,284 | 35,424 |
| `log/oneline-100@git-git` | 4.76 | 6.26 | 6.15 | 21,986 | 20,188 |
| `log/oneline-100@synth-100k` | 3.99 | 5.17 | 5.14 | 16,512 | 15,604 |
| `log/path@git-git` | 366.65 | 450.93 | 446.27 | 244,442 | 243,682 |
| `log/stat-n50@git-git` | 102.16 | 392.18 | 389.08 | 113,432 | 110,746 |
| `show/head@git-git` | 3.37 | 4.67 | 4.66 | 16,256 | 15,504 |
| `blame/file@git-git` | 313.05 | 308.04 | 303.46 | 234,546 | 233,630 |
| `revs/count@git-git` | 540.46 | 767.01 | 756.98 | 126,158 | 126,128 |
| `cat/blob@git-git` | 2.73 | 4.03 | 4.05 | 13,308 | 13,482 |
| `status/clean@synth-100k` | 124.51 | 105.37 | 104.76 | 34,134 | 34,388 |

Raw samples and argv are in
[`log-index-4729-paired.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/log-index-4729-paired.json).
The standard daily captures are `daily-log-index-4729-before.json` and
`daily-log-index-4729-after.json`; the latter also includes the runner's paired
headline baseline. Its extra baseline has three samples, so use the 20-run table
and 300-run Hyperfine check above for the before/after comparison.

A warm loop over the same 100 real commit ids, repeated 10,000 times, measured
about **78 → 24 ns per lookup**. The standalone Criterion target covers a broader
4,096-id scattered working set, plus hot hit/miss controls, without preparing the
objects suite's unrelated pack fixtures. On the real index, the fixed scattered
hit/miss batches measured 260/289 µs per 4,096 lookups (about 64/71 ns per lookup).

```bash
# Default: deterministic 200k-object index.
cargo bench -p grit-lib --bench pack_index
# Real repository: choose one of its .idx files.
GRIT_PACK_INDEX_BENCH=/path/to/repo/.git/objects/pack/pack-HASH.idx \
  cargo bench -p grit-lib --features http-ureq,tracing --bench pack_index
```

## Bounded log and loose clone tags (2026-10-27)

Issue [#2080](https://github.com/gitbutlerapp/grit/issues/2080) exposed pack mmap
residency from scattered tag and terminal-object probes. Whole-object info now uses
bounded positioned reads; small whole tags use the same sparse read path in both
general and single-pack readers. Delta, oversized, and malformed entries retain
the existing readers. Clone packs imported tags and remote-tracking refs with the
shared ref writer and CAS pruning, while retaining symbolic refs and reflogs.
Directory cleanup is batched, and lock configuration is loaded once per pack.

Sparse-read descriptors are owned by a repository-scoped pool capped at **16**,
independently of the number of cached pack mappings. Evicted packs continue through
the existing mapped reader without reopening paths, preserving the mapped inode
when a pack is replaced. In-flight positioned reads keep their descriptor until
completion. If opening a new pack hits the OS descriptor limit, the pool is
cleared before retrying, leaving cached mappings valid. A Linux subprocess
regression reads 96 mapped packs twice with `RLIMIT_NOFILE=64`, also testing a
host that has reserved 48 descriptors. It checks all 96 tag decorations and
compares the CLI's bounded JSON output with an unrestricted run. The scattered-read RSS and path-replacement
regressions cover retained descriptors and eviction separately.

Measured on a 4-vCPU AMD EPYC 7763 Linux VM with 16 GiB RAM, Ubuntu 24.04.5,
Rust 1.99.0 `release` builds, and stock Git 2.55.0. The full git/git source is
`37b3475f`, with 1,012 tags. Frozen main is `d99b149b`; the implementation is
`e9dc7444`. Each log cell is the median of nine rotated, interleaved fresh
processes after one warmup per tool, with builds and tests idle. Python
`perf_counter` wraps `/usr/bin/time -v`; memory is median process peak RSS in KiB.
Git uses the issue's full author/committer/parent/decoration/message format, and
grit uses `log -n 10 --json`. Both representations have identical objects and refs;
the loose state has no cached packed tag peel records.

| Ref representation | Git ms | Main ms | Fixed ms | Git RSS | Main RSS | Fixed RSS |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| Loose tags | 21.77 | 37.26 | 38.08 | 16,772 | 45,556 | 24,124 |
| Packed tags | 14.00 | 5.46 | 5.33 | 16,736 | 16,980 | 16,412 |

Loose-tag RSS falls **47.0%**, to **1.44× Git**, and the representation gap falls
from 2.68× to 1.47×. The loose-tag median is 0.82 ms higher (2.2%); the packed
median is slightly lower. The Linux regression uses a deterministic 96 MiB pack
with 768 scattered annotated tags and checks equal bounded JSON output before and
after packing. Its loose-only increase falls from **48,940 KiB to 452 KiB**.

Clone measurements use one warmup and three rotated fresh local clones per tool.
Ref listings match system Git, and the fixed clone passes `git fsck --strict`
(with the historical upstream warnings). Clone memory is not the destination of
the log savings:

| Local clone | ms | Peak RSS KiB | Loose tags afterward |
| --- | ---: | ---: | ---: |
| Git | 461.56 | 87,484 | 0 |
| Main | 661.15 | 111,344 | 1,012 |
| Fixed | 872.39 | 113,556 | 0 |

Safe tag packing adds about **0.21 s** to this local clone (32%), with about
**2.2 MiB** more clone RSS. The resulting daily log uses the packed row above.
Avoiding the temporary loose tag writes during initial ref import remains a
possible clone optimization; this change retains the existing lock/CAS semantics.
Existing loose-tag repositories benefit from the sparse reads immediately.

A packed-ref 1,000-commit metadata control checks the shared object-read path:
main **23.28 ms / 29,208 KiB**, fixed
**23.58 ms / 28,444 KiB**, and Git
**27.68 ms / 18,596 KiB** (the same nine-sample method).

The daily matrix tracks `log/json-n10-loose-tags@git-git` and
`log/json-n10-packed-tags@git-git`. The committed before/after daily baselines use
maximum sampled RSS for their schema's `peak_rss_bytes`; the table above uses the
median required by the issue. All raw samples, initial reproduction, clone
measurements, and fixture details are in
[`log-loose-tags-2080-study.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/log-loose-tags-2080-study.json).

### Descriptor-pressure follow-up (2026-10-27)

The bounded descriptor pool was checked again on the same VM shape and git/git
revision (`37b3475f`). Frozen `origin/main` is `e98de8a163`; reviewed is
`59557cb17` applied over that workspace target. The fixed build adds descriptor
pooling and exhaustion recovery. The compared crates were rebuilt between source
snapshots, and the raw study records distinct binary hashes. Remaining verification
work was paused during measurement. Log cells are nine-sample interleaved medians
after one warmup per tool; clone cells use one warmup and three rotated samples.
Every cell is **wall ms / peak RSS KiB**:

| Scenario | Git | origin/main | Reviewed | Fixed |
| --- | ---: | ---: | ---: | ---: |
| Log 10, loose tags | 21.24 / 16,816 | 36.62 / 44,836 | 37.55 / 25,664 | 37.62 / 25,908 |
| Log 10, packed tags | 14.33 / 16,760 | 5.30 / 17,020 | 4.87 / 17,476 | 4.63 / 17,784 |
| Log 1,000, packed tags | 27.80 / 18,656 | 23.07 / 29,988 | 22.97 / 30,052 | 22.67 / 30,236 |
| Log 96, 96 packs | 7.45 / 6,852 | 17.08 / 16,128 | 16.93 / 16,008 | 16.73 / 16,380 |
| Local clone | 470.13 / 103,460 | 675.95 / 111,500 | 858.88 / 112,620 | 854.66 / 112,784 |

The final loose-tag log uses **1.54× Git's RSS**, versus **2.67×** on main.
Descriptor bounding adds **0.08 ms** and **244 KiB** to the reviewed loose-tag
median; packed log and local clone medians are slightly lower. The clone layout
remains unchanged: main leaves 1,012 loose tags; reviewed, fixed, and Git leave none.

With `RLIMIT_NOFILE=64`, Git, main, and fixed all return 96 commits from the
96-pack fixture. Reviewed fails with an I/O error after exhausting descriptors.
The library regression also checks every tag decoration and recovery when a host
has already reserved 48 descriptors. The CLI RSS fixture's loose-versus-packed
increase is **628 KiB**, within its 12 MiB allowance. Raw samples and controls:
[`log-loose-tags-2080-descriptors.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/log-loose-tags-2080-descriptors.json).

## Daily regression audit and deep-tree ancestor checks (2026-10-28, #5019)

The October 27 daily report flagged seven scenarios whose grit/git ratio rose
against October 26. Grit's own medians mostly *fell* between those days; the
ratios rose because the October 26 sitting ran Git 1.5–2.6× slower than both
neighbouring days (`diff/worktree@deep-tree` Git: 100.8 → 244.8 → 92.4 ms on
October 25, 26 and 27). The two reported revisions (`c7cf2c4436`, `8dbc58e772`)
and main (`307186978`) were built in `release` and run with `grit-bench daily`
against stock Git 2.43.0 in one sitting on a 4-vCPU Intel Xeon VM with 16 GiB RAM.
None of the seven reproduces as a code regression:

| Scenario | Git ms (new run) | c7cf2c ms | 8dbc58 ms | main ms | Grit RSS c7cf2c → 8dbc58 (KiB) |
| --- | ---: | ---: | ---: | ---: | ---: |
| `commit/staged@synth-100k` | 86.83 | 79.70 | 83.57 | 82.20 | 44,804 → 44,920 |
| `diff/staged@deep-tree` | 46.43 | 48.89 | 49.45 | 44.42 | 49,044 → 48,856 |
| `diff/worktree@deep-tree` | 118.98 | 278.18 | 259.95 | 256.55 | 71,740 → 71,568 |
| `reset/hard@synth-10k` | 21.35 | 28.76 | 31.18 | 28.28 | 24,864 → 24,512 |
| `switch/branch@synth-100k` | 184.19 | 119.58 | 100.90 | 91.30 | 42,940 → 43,300 |
| `switch/delete-heavy@wide-48k` | 142.97 | 162.16 | 146.66 | 144.01 | 33,352 → 34,044 |
| `switch/large-blob@big-asset` | 949.32 | 641.14 | 631.06 | 795.69 | 18,312 → 18,952 |

`switch/large-blob@big-asset` writes a 300 MB file twice per sample and swings
between roughly 480 and 840 ms from one sitting to the next. A ten-run
`hyperfine` on one copied fixture put all three builds within noise
(c7cf2c 858 ± 49 ms, 8dbc58 822 ± 24 ms, main 845 ± 17 ms; Git 1,510 ms).

The standing gap was `diff/worktree@deep-tree`, about 2.2× Git. `strace -c`
showed grit and Git making the same 182,000 `stat` calls: 5,000 files 31–40
levels deep have 175,000 distinct ancestor directories, and both tools check
every ancestor for a symlink before trusting a file's stat. Grit spent another
250 ms of CPU in the cache that remembered those answers, a single
mutex-guarded map shared by the four stat workers. Each entry needed around 35
contended lock round-trips and an allocated key per ancestor.

Two library changes remove that cost:

1. Each scan worker owns its ancestor cache. It remembers the last chain of
   ancestors verified to be real directories and the first one that is not, so
   sorted index paths only probe the components that differ from the previous
   path. There is no locking and no per-directory allocation. Staging no longer
   repeats the same check against a second locked cache unless case folding or
   Unicode precomposition can make the two spellings differ.
2. On 64-bit Linux, when two or more components are still unverified, one
   `openat2(..., RESOLVE_NO_SYMLINKS)` call relative to a handle on the work tree
   confirms the whole directory chain in a single path walk. Separate `lstat`
   calls re-walk the path from the root, which is quadratic in depth. Any failure,
   whether a symlink, a missing path, a file, or a kernel without `openat2`, falls
   back to the per-component probe, which finds and caches the exact ancestor.

On the copied deep-tree fixture (20 runs, one prepare per run), `grit diff` went
from 240.7 ± 23.8 ms to 85.3 ± 5.2 ms against Git's 121.8 ± 8.2 ms. User CPU fell
from 269 to 45 ms, system CPU from 359 to 130 ms, and syscalls from 182,000 `statx`
to 5,000 `openat2` plus 7,000 `statx`. Output is byte-identical. The same
binaries on related daily scenarios, main then fixed in one sitting:

| Scenario | Git ms | Grit main ms | Grit fixed ms | Main RSS KiB | Fixed RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: |
| `add/all-ignorecase@synth-100k` | 169.51 | 239.44 | 223.91 | 66,712 | 66,608 |
| `add/all@synth-100k` | 171.91 | 161.20 | 163.17 | 48,436 | 49,000 |
| `add/modified@synth-10k` | 21.55 | 13.05 | 16.57 | 26,300 | 26,288 |
| `add/update@synth-100k` | 103.89 | 112.41 | 100.05 | 44,516 | 44,148 |
| `commit/small@synth-100k` | 1005.22 | 166.57 | 169.77 | 48,876 | 48,832 |
| `commit/small@synth-10k` | 33.63 | 22.07 | 25.93 | 25,500 | 25,596 |
| `commit/staged@synth-100k` | 80.10 | 91.47 | 83.25 | 45,816 | 45,924 |
| `diff/staged@deep-tree` | 44.15 | 46.13 | 45.73 | 49,940 | 50,148 |
| `diff/staged@synth-100k` | 73.13 | 45.52 | 43.60 | 69,032 | 69,412 |
| `diff/stat@synth-100k` | 110.34 | 78.27 | 73.49 | 58,852 | 59,124 |
| `diff/worktree@deep-tree` | 119.75 | 256.94 | 78.87 | 72,704 | 47,364 |
| `diff/worktree@synth-100k` | 122.03 | 83.90 | 77.78 | 68,828 | 69,260 |
| `diff/worktree@synth-10k` | 7.39 | 7.83 | 8.40 | 28,348 | 28,548 |
| `reset/hard-delete-heavy@wide-48k` | 234.45 | 281.33 | 250.40 | 52,596 | 53,368 |
| `reset/hard@synth-10k` | 15.53 | 30.97 | 26.27 | 25,560 | 26,024 |
| `reset/mixed@synth-10k` | 8.20 | 12.05 | 11.05 | 20,944 | 24,028 |
| `reset/sparse@sparse-120k` | 28.60 | 26.98 | 25.63 | 20,964 | 21,444 |
| `restore/paths@synth-10k` | 11.92 | 15.36 | 13.11 | 23,652 | 23,936 |
| `stash/push-pop@synth-10k` | 252.14 | 155.67 | 163.65 | 26,080 | 26,688 |
| `status/clean@deep-tree` | 1300.18 | 747.81 | 612.02 | 72,216 | 45,632 |
| `status/clean@synth-100k` | 79.59 | 72.37 | 72.42 | 37,676 | 38,252 |
| `status/dirty@deep-tree` | 1400.01 | 769.86 | 693.88 | 79,148 | 56,616 |
| `status/dirty@synth-100k` | 121.12 | 66.03 | 80.54 | 42,796 | 43,232 |
| `status/dirty@synth-10k` | 12.81 | 10.61 | 11.92 | 20,556 | 21,292 |
| `switch/branch@synth-100k` | 155.51 | 102.50 | 89.20 | 44,180 | 44,312 |
| `switch/delete-heavy@wide-48k` | 125.65 | 158.45 | 132.89 | 34,888 | 34,884 |
| `switch/large-blob@big-asset` | 860.87 | 608.82 | 774.67 | 19,504 | 19,696 |
| `switch/sparse@sparse-120k` | 42.51 | 25.95 | 25.30 | 21,016 | 21,560 |

`diff/worktree@deep-tree` is **69% faster** and now beats Git, with peak RSS down
from 71 to 46 MiB. Deep-tree status gains 10–18% and 22–31 MiB. The shallow
fixtures have one or two directory levels, so the change barely touches them.
Their slower-looking rows were re-run in four alternating batches (fixed, main,
fixed, main). `status/dirty@synth-100k` came out at 68.9 / 65.9 / 61.8 / 64.6 ms,
`add/modified@synth-10k` at 13.0 / 11.9 / 12.3 / 11.8 ms,
`commit/small@synth-10k` at 21.2 / 21.7 / 20.3 / 20.4 ms, and `switch/large-blob`
at 634 / 842 / 798 / 795 ms. A 150-run order-swapped `hyperfine` of `grit status`
on synth-10k measured 10.0 and 9.6 ms fixed against 9.5 and 10.4 ms main.
These rows are noise.

Raw reports: [c7cf2c](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-5019-old.json),
[8dbc58](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-5019-new.json),
[main, reported scenarios](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-5019-main.json),
[main](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-5019-before.json),
[fixed](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-5019-after.json),
and the alternating confirmation batches
([1](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-5019-confirm-1-after.json),
[2](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-5019-confirm-2-before.json),
[3](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-5019-confirm-3-after.json),
[4](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-5019-confirm-4-before.json)).
Library tests compare the per-worker cache with an uncached probe for every
query order, including a symlink at the top, in the middle, and as the last
directory, as well as missing ancestors and file ancestors. They also cover a work
tree reached through a symlink and the `openat2` check on its own.

## Daily regression audit (2026-10-27, #4744)

The original daily reports used different hardware and Git versions: October 25
used an Intel Xeon with Git 2.43.0, while October 26 used an AMD EPYC with Git 2.55.0.
To separate code regressions from environment changes, both reported revisions
(`2824ad3911`, `c7cf2c4436`) and frozen main `80dc322f7` were built in `release` and
measured against stock Git 2.55.0 on one 4-vCPU AMD EPYC 7763 Linux VM with 16 GiB RAM.
No compilation or tests ran during native timing. Historical and final paired comparisons use ten
samples per tool and one warmup, with the frozen daily harness capped at ten.
The three repaired plans use the corrected harness with `--min-runs 10` and its
normal convergence policy (up to twenty runs). Tables show median elapsed
milliseconds and maximum child-process RSS in KiB. The daily runner uses wait4;
separate `/usr/bin/time -v` runs corroborate the three changed operations.

HTTP measurements keep the c7cf2c4436 `grit-http-server` binary constant for both
clients. The git/git fixture is pinned to v2.47.0 (`777489f9e0`). Preparation,
commands, every sample, and phase measurements are retained in the linked reports.
The harness is identical within each final main/fixed pair. The repaired runner honors
`--min-runs 10`, rejects placeholder `true` plans, supplies Git's editor for
history reword, derives an actual change-id prefix, and starts deepen with a
single selected-branch shallow boundary. Thus repaired scenarios have a valid
main/fixed control, but cannot be compared to the old failed/no-op daily series.

The fixes preserve the features introduced by the regression commits. Streaming
restaging now hashes and stores once when local canonical blob size proves that
the content changed; unchanged, missing, and same-size objects retain the
hash-first path, without fetching a missing previous blob. Path commits validate
pathspecs without staging a discarded index, retaining both real index views and
hook behavior. Push maintains an incremental count of interesting queued commits,
retaining the existing reachability cutoff while removing repeated heap scans.
Callgrind attributed 54.6% of the force-with-lease run's instructions to those
scans. `perf` was unavailable because the VM disallows hardware performance events.

The commit-boundary tests below were built from fresh archived source trees,
with source mtimes refreshed before each build. Cached-build samples that failed
this check were discarded. These are separate ten-run experiments.

| Scenario | Parent | Introducing commit / merge | Git ms (after) | Grit parent ms | Grit after ms |
| --- | --- | --- | ---: | ---: | ---: |
| `add/modified@big-blobs` | `818de8006e` | af8594f21 | 291.32 | 137.40 | 237.29 |
| `commit/paths@synth-10k` | `c0b0c50bfa` | bebfaa546 via 2a0f003c1 | 40.79 | 51.35 | 59.35 |
| `push/force-with-lease@git-git` | `6388d987e4` | 69d7a1fa6 via 741ad80ab | 159.63 | 691.83 | 1456.57 |
| `push/incremental@git-git` | `6388d987e4` | 69d7a1fa6 via 741ad80ab | 184.11 | 1312.61 | 2066.00 |
| `push/incremental-http@git-git` | `6388d987e4` | 69d7a1fa6 via 741ad80ab | 99.52 | 1249.60 | 2017.09 |

### Historical reproduction

Each row compares the two reported commits on this VM. RSS is the maximum
observed grit child-process RSS; the Git time comes from the newer-commit run.

| Scenario | Git ms | 2824ad ms | c7cf2c ms | Grit RSS old → new (KiB) |
| --- | ---: | ---: | ---: | ---: |
| `add/modified@big-blobs` | 287.61 | 138.46 | 239.14 | 86,868 → 88,976 |
| `branch/delete@many-refs` | 2.46 | 9.94 | 9.46 | 15,828 → 16,712 |
| `clone/bare@synth-10k` | 701.23 | 775.77 | 780.35 | 23,712 → 24,624 |
| `commit/paths@synth-10k` | 38.82 | 51.65 | 58.71 | 28,576 → 30,180 |
| `commit/small@synth-100k` | 939.51 | 258.01 | 256.64 | 61,060 → 62,168 |
| `commit/small@synth-10k` | 38.91 | 43.24 | 35.97 | 26,792 → 28,784 |
| `commit/staged@synth-100k` | 110.34 | 77.60 | 75.05 | 57,272 → 58,900 |
| `fetch/dry-run@git-git` | 111.82 | 99.24 | 97.93 | 92,376 → 99,932 |
| `fetch/incremental-http@git-git` | 1957.74 | 1418.89 | 1433.68 | 100,064 → 102,296 |
| `fetch/incremental@git-git` | 113.35 | 152.33 | 146.69 | 94,004 → 92,736 |
| `fetch/one-branch@git-git` | 114.04 | 152.62 | 149.94 | 92,400 → 94,824 |
| `pull/check@git-git` | 115.63 | 209.64 | 154.70 | 93,888 → 93,136 |
| `pull/fast-forward@git-git` | 131.72 | 216.22 | 157.08 | 92,424 → 93,252 |
| `push/force-with-lease@git-git` | 159.01 | 693.04 | 1477.63 | 74,092 → 75,408 |
| `push/incremental-http@git-git` | 95.90 | 1226.90 | 1967.57 | 192,168 → 193,300 |
| `push/incremental@git-git` | 179.50 | 1310.88 | 2091.71 | 192,316 → 192,840 |
| `remote/remove@many-refs` | 622.55 | 779.77 | 774.16 | 25,004 → 25,836 |
| `status/paths@synth-100k` | 44.60 | 109.47 | 82.89 | 41,972 → 43,512 |
| `switch/branch@synth-100k` | 239.07 | 106.01 | 105.22 | 53,380 → 54,908 |

Only large-blob add, path commits, and the three push cases reproduce a
meaningful slowdown. Branch delete, bare clone, other commits, fetch/pull, remote
remove, path status, and switch memory do not reproduce the daily regression;
no speculative fixes were made for them. Main also already includes the MIDX
residency improvement, which accounts for its lower network memory versus c7cf2c.

The old deepen fixture produced over a thousand shallow boundaries; stock Git
was still negotiating after 204 seconds and was deliberately terminated. That
exit is not treated as a reproduction of the reported code 128. Reword fails
with Git 2.55 code 129 because the old plan supplied an unsupported message
option. Change-id resolution timed `true` in both old plans and is excluded
from timing reports. The corrected plans test actual operations and compare
Git/grit results, including exact shallow boundary ids.

### Main versus fixed, including related commands

The fixed binaries contain the library and harness changes through `66111b8f2`;
the final branch additionally strengthens a test fixture and records this audit.
Git ms is the fixed-run control; both Git controls are retained in the raw JSON.
The `commit/small@synth-10k` row includes the initial ten-run batch and three
thirty-run confirmation batches, with batch order alternated: **100 samples per
binary**, without dropping outliers. Its initial 34.85 → 45.98 ms result was
bimodal in both Git and grit. Repeated grit medians were 34.4–34.5 ms on main and
33.3–34.0 ms fixed, so no persistent slowdown reproduced.

| Scenario | Git ms | Grit main ms | Grit fixed ms | Main RSS KiB | Fixed RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: |
| `add/all@synth-100k` | 241.57 | 232.62 | 233.59 | 49,748 | 47,216 |
| `add/dry-run@synth-10k` | 20.14 | 23.02 | 22.44 | 20,812 | 21,248 |
| `add/modified@big-blobs` | 289.87 | 238.70 | 137.63 | 88,476 | 89,584 |
| `add/modified@synth-10k` | 19.44 | 19.48 | 18.75 | 24,284 | 24,476 |
| `add/patch-file@synth-10k` | 8.38 | 11.08 | 10.37 | 19,776 | 19,880 |
| `add/update@synth-100k` | 153.38 | 160.58 | 158.52 | 44,740 | 44,648 |
| `branch/delete@many-refs` | 2.46 | 10.59 | 9.89 | 16,700 | 16,436 |
| `clone/bare@synth-10k` | 704.71 | 784.56 | 784.58 | 24,572 | 24,668 |
| `commit/dry-run@synth-10k` | 23.27 | 49.85 | 48.05 | 23,700 | 23,944 |
| `commit/paths@synth-10k` | 40.54 | 57.11 | 51.20 | 26,672 | 27,180 |
| `commit/small@small` | 9.73 | 10.85 | 9.94 | 18,580 | 19,216 |
| `commit/small@synth-100k` | 934.13 | 244.88 | 246.85 | 51,408 | 51,996 |
| `commit/small@synth-10k` | 39.11 | 34.45 | 33.72 | 25,824 | 26,360 |
| `commit/staged@synth-100k` | 108.32 | 68.59 | 67.53 | 47,948 | 48,860 |
| `fetch/deepen@git-git` | 2666.77 | 8128.51 | 8163.79 | 136,208 | 134,960 |
| `fetch/dry-run@git-git` | 113.73 | 75.95 | 74.65 | 64,024 | 64,024 |
| `fetch/incremental-http@git-git` | 1950.46 | 1378.59 | 1349.62 | 69,212 | 69,368 |
| `fetch/incremental@git-git` | 114.33 | 78.11 | 76.35 | 63,732 | 63,876 |
| `fetch/one-branch@git-git` | 113.58 | 78.03 | 76.94 | 64,104 | 64,208 |
| `history/reword@synth-10k` | 137.11 | 261.97 | 259.87 | 23,948 | 24,404 |
| `pull/check@git-git` | 117.48 | 77.69 | 76.82 | 64,056 | 64,116 |
| `pull/fast-forward@git-git` | 132.99 | 82.16 | 80.69 | 64,144 | 64,288 |
| `push/force-with-lease@git-git` | 162.01 | 1453.69 | 749.27 | 75,352 | 75,684 |
| `push/incremental-http@git-git` | 93.89 | 2022.37 | 1296.89 | 107,168 | 108,304 |
| `push/incremental@git-git` | 185.40 | 2113.62 | 1368.53 | 106,724 | 107,784 |
| `remote/remove@many-refs` | 613.84 | 771.89 | 775.66 | 25,616 | 25,652 |
| `resolve/change-id@git-git` | 663.65 | 748.31 | 726.20 | 113,548 | 113,972 |
| `status/paths@synth-100k` | 44.20 | 81.24 | 81.88 | 41,408 | 43,104 |
| `switch/branch@synth-100k` | 235.74 | 85.09 | 82.64 | 45,740 | 46,164 |
| `add/hunks@large-file` | unavailable | unavailable | unavailable | unavailable | unavailable |
| `add/lines@large-file` | unavailable | unavailable | unavailable | unavailable | unavailable |
| `add/patch-file@large-file` | unavailable | unavailable | unavailable | unavailable | unavailable |
| `push/refspec@git-git` | unavailable | unavailable | unavailable | unavailable | unavailable |

The four unavailable related plans fail on both binaries: large-file hunk and
patch-file plans fail in stock Git with code 128; the line plan fails preparation
with code 1; the refspec plan's correctness setup rejects a non-fast-forward push.
They have no publishable timing or RSS series and are left for separate repairs.
The historical and main harness code for these plans is identical.

Large-blob add is **42.3% faster**, path commits **10.4% faster**, force-with-lease
push **48.5% faster**, HTTP incremental push **35.9% faster**, and local incremental
push **35.3% faster**. These are elapsed-time improvements with essentially unchanged
memory on the changed operations; the original large memory increases did not
reproduce. Same-size streams deliberately retain hash-first behavior so that a
stat refresh does not write unchanged objects. Push remains slower than Git;
corrected deepen, reword, and change-id plans now expose valid baselines for
future work rather than substituting numbers for the old failed/no-op series.

Separate GNU time runs retain every matching CLI-operation record, including
warmups, tracing, and correctness checks. These measure the CLI child directly;
the primary daily helper also includes shell/pipeline launch overhead. GNU time's
wall time is quantized to 10 ms, so the small path-commit change is resolved by
the primary measurements rather than the rounded values below. RSS is the maximum
observed child peak, in KiB; complete GNU time output is retained in the audit.

| GNU time confirmation | Git ms | Grit before ms | Grit after ms | Before RSS KiB | After RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: |
| `add/modified@big-blobs` | 280 | 230 | 130 | 84,988 | 85,352 |
| `commit/paths@synth-10k` | 30 | 50 | 50 | 22,712 | 24,272 |
| `push/force-with-lease@git-git` | 160 | 1440 | 750 | 73,836 | 73,928 |
| `push/incremental-http@git-git` | 90 | 1970 | 1250 | 104,868 | 105,960 |
| `push/incremental@git-git` | 180 | 2040 | 1330 | 104,640 | 105,872 |

Raw reports: [historical old](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-4744-old.json),
[historical new](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-4744-new.json),
[main](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-4744-before.json),
[fixed](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-perf-4744-after.json),
and [audit, commit boundaries, confirmations, and GNU time output](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/perf-4744-audit.json).
The audit also records binary hashes and Criterion samples for
`send_pack/backdated_want_1000_haves` (17.72 ms after, ten samples).

Library regressions cover one-pass size-changed streaming, stale stat sizes,
missing objects without a promisor fetch, validation without hashing, two real
staging passes per path commit, and duplicate/shared want-have frontiers. A new
system-Git round trip covers SHA-1/SHA-256 and binary CRLF attributes at two large
file thresholds. Existing scoped-commit/hook, pack/fsck, push flag/lease, and the
17 HTTP promisor tests pass. The optional external deep-history guard returns
early without its separate fixture; the new frontier benchmark is self-contained.

Validation remains limited by pre-existing failures. The required full
`cargo test -p grit-lib -p grit-cli` command fails to compile
`history_injected_ref_store` (`Signature: From<&str>`); the library-only full
command also exposes stale `ServeOptions` calls in `serve_upload_pack_inprocess`.
CLI units report 155 passes and the same command-page ordering failure before
and after. Library units initially report 2,628 passes, two failures, and three
ignored; the post-change result and exact failures are recorded in the audit.
The reset/fsck failure observed after these changes also reproduces in isolation
on freshly archived, unmodified main. No tests were removed or weakened.
Workspace clippy, formatting, rustdoc with warnings denied, and the docs content
check pass. The benchmark utility reports 132 passing units; smoke tests exercise
the repaired plans and a non-smoke test verifies a requested thirty-run minimum.

Reproduce the five improved scenarios with each frozen release binary and the
same stock Git, keeping compilation and tests stopped during measurement:

```bash
cargo build --release -p grit-cli -p grit-utils -p grit-http-server
for scenario in add/modified@big-blobs commit/paths@synth-10k \
  push/force-with-lease@git-git push/incremental-http@git-git \
  push/incremental@git-git; do
  grit-bench --grit /path/to/frozen/grit --min-runs 10 daily \
    --only "$scenario" --output "${scenario//\//_}.json"
done
cargo bench -p grit-lib --bench pack_frontier --features http-ureq,tracing
```

For HTTP comparisons, put the same frozen `grit-http-server` first on `PATH` for
both clients. Historical ten-run reports used a frozen c7cf2c4436 harness with
its maximum capped at ten; the repaired plans and thirty-run confirmations use
the corrected runner's `--min-runs` option.

## Branch rename with many refs (2026-10-27)

Renaming a loose branch used to build a complete packed-ref map only to validate
it before removing the old name. Ref removal now scans borrowed names through the
same strict parser and avoids constructing replacement contents when no packed
record matches. Validation still checks the entire file, including malformed
records after the last matching name. Reflog, locking and worktree behavior are
unchanged.

On a 4-vCPU AMD EPYC 9V74 Linux VM with 16 GiB RAM, Rust 1.99.0 `release` builds
and system Git 2.55.0, the full daily fixture has 1,000 commits and 10,000 refs.
The baseline is `d989c8fe8`; the after build includes the branch-rename optimization.
The rename row calls the `rename_branch` library API through
`grit-bench rename-branch-lib`; the other rows exercise `grit` commands.
Compilation and test processes were paused for measurement. Each daily row has
one warmup and 20 runs per tool; times are medians, RSS is the maximum reported by
the daily measurement helper. The Git columns show before / after measurements
in the same sitting.

| Scenario | Git ms (before / after) | Grit before ms | Grit after ms | Grit RSS KiB (before / after) |
| --- | ---: | ---: | ---: | ---: |
| `branch/contains-json@git-git-10k-refs` | 36.30 / 36.54 | 31.48 | 30.83 | 35440 / 35064 |
| `branch/contains@git-git-10k-refs` | 35.94 / 36.10 | 30.44 | 30.54 | 35284 / 35172 |
| `branch/create@many-refs` | 1.87 / 1.83 | 3.47 | 3.57 | 14856 / 13348 |
| `branch/delete@many-refs` | 2.08 / 2.08 | 8.21 | 5.09 | 15824 / 14272 |
| `branch/list-json@many-refs` | 2.00 / 1.99 | 3.16 | 3.15 | 14392 / 14928 |
| `branch/list@many-refs` | 1.69 / 1.70 | 2.83 | 2.85 | 13456 / 13196 |
| `branch/list@many-refs-loose` | 1.74 / 1.67 | 2.96 | 2.91 | 13640 / 13300 |
| `branch/merged-ref@many-refs` | 1.17 / 1.23 | 2.65 | 2.69 | 10596 / 10320 |
| `branch/merged@many-refs` | 1.85 / 1.86 | 3.17 | 3.18 | 15008 / 15140 |
| `branch/rename@many-refs` | 2.06 / 2.05 | 6.99 | 4.48 | 12104 / 10368 |
| `tag/create@many-refs` | 1.82 / 1.81 | 3.60 | 3.61 | 14884 / 14684 |
| `tag/list@many-refs` | 4.20 / 4.09 | 5.41 | 5.08 | 14040 / 13724 |

A separate 40-run interleaved direct-spawn comparison confirmed rename at
**6.85 → 4.40 ms**, against Git's **1.73 ms**. Ten `/usr/bin/time -v` samples per
binary measured peak RSS at **12,112 → 10,384 KiB** (Git: **5,128 KiB**).
Heaptrack counted **25,141 → 5,060 allocations** and **2.72 → 1.27 MB** peak live
heap; these heap measurements exclude the profiler's own memory.

Small changes in the daily create/list rows were checked with 100 interleaved
runs per binary: create **2.88 → 2.90 ms**, list **2.41 → 2.42 ms**, JSON list
**2.70 → 2.71 ms**. These differences are under 1%; the repeated daily create
samples varied more, so no improvement is claimed for these paths.

Raw daily samples are in
[`daily-branch-rename-many-refs-before.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-branch-rename-many-refs-before.json)
and
[`daily-branch-rename-many-refs-after.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-branch-rename-many-refs-after.json).
The interleaved runs, GNU time RSS samples, Heaptrack totals and daily repeats
are in
[`branch-rename-many-refs-study.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/branch-rename-many-refs-study.json).

## Multi-pack-index preparation and clean status (2026-10-26)

Issue [#2064](https://github.com/gitbutlerapp/grit/issues/2064) exposed three MIDX
mappings and repeated checksum scans during a clean status check, plus a whole-file
allocation just to read pack names. MIDX preparation and reads now use the owning
repository's `PackStore`. Strict and permissive views share validated layer backing
while retaining their own chain acceptance and diagnostic behavior. Checksums remain
mandatory, with a 2 MiB residency budget and checksum-only pages released afterward.
Pack-name reads reuse prepared metadata or fault in only the relevant mapped pages.

Measured on a 4-vCPU AMD EPYC 7763 Linux VM with 16 GiB RAM, Rust 1.99.0 `release`
builds and stock Git 2.55.0. The exact issue fixture is git/git at `7e34d817`, with
82,563 reachable commits, 422,320 packed objects, packed refs, a changed-path commit
graph, and a Git-written 12 MB MIDX. Frozen `origin/main` (`dbdf37d29`) and fixed
binaries were run in one sitting with builds and tests paused. Each cell is the
median of seven rotated warm-cache runs after three warmups. Python `perf_counter`
wraps `/usr/bin/time -v`, including process launch; RSS is in KiB.

| Status / MIDX | Git ms | Main ms | Fixed ms | Git RSS | Main RSS | Fixed RSS |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| human / present | 13.99 | 45.28 | 24.56 | 7,664 | 61,912 | 17,588 |
| short / present | 13.48 | 44.00 | 24.35 | 7,580 | 61,912 | 17,848 |
| JSON / present | 13.59 | 44.02 | 24.45 | 7,704 | 61,884 | 17,716 |
| Markdown / present | 13.82 | 44.07 | 24.36 | 7,696 | 61,780 | 17,892 |
| human / absent | 13.21 | 15.78 | 15.89 | 7,580 | 16,340 | 17,084 |

Git's JSON and Markdown controls use ordinary `git status`; the short control uses
`git status --short`. All present/absent samples for every format are committed in
`grit-utils/baselines/midx-status-2064.json`. Human status removes **98.9% of the
MIDX-only RSS increase** (45,572 → 504 KiB), reduces total RSS by **71.6%**, and
reduces elapsed time by **45.8%**. Without MIDX, elapsed time remains near 15.8 ms;
the measured release-binary RSS increases by 744 KiB. The remaining base footprint
and fresh-process checksum cost leave room for further work against Git.

Criterion separates fresh repository-cache preparation from warm lookups (20
samples, 1-second warmup, 2-second measurement, OS caches warm). Preparing strict
and permissive views over the real 12 MB MIDX improves **14.69 → 7.36 ms**. Warm
lookups on the existing ten-pack fixture improve **96.99 → 59.27 ns** for hits and
**82.42 → 47.26 ns** for misses. The unchanged all-pack-index hit control measures
29.63 → 29.62 ns; its tiny miss control measures 5.32 → 6.23 ns.

```bash
GRIT_MIDX_BENCH_OBJECTS=/path/to/git/.git/objects \
  cargo bench -p grit-lib --bench midx -- midx_prepare
cargo bench -p grit-lib --bench midx -- midx_lookup
cargo test -p grit-cli --test status_midx_memory -- --nocapture
```

The Linux RSS regression uses 420,000 historical blobs and a one-file clean
checkout, exercising human, short, JSON, and Markdown output with MIDX present and
absent. Library regressions cover one mapping per owner, explicit and implicit
caller contexts, concurrent views, independent stores, SHA-1/SHA-256, refresh,
missing/malformed layers, diagnostics, corruption fallback, and system-Git fsck
round trips.

## PIE startup relocation probe

Current policy: installers keep the stock GNU asset. Fresh final-tip source
measurements below reproduce a RELR dist slowdown; the optional asset remains
manual-only. Earlier results retain their original revisions and hosts.

Issue [#2028](https://github.com/gitbutlerapp/grit/issues/2028) investigates loader
work without disabling PIE or ASLR. A CLI-only `-z pack-relative-relocs` probe at
`f0e7cddd5` keeps both Cargo profiles, dependency features and command behavior
unchanged. On a 4 vCPU AMD EPYC 9V74 Linux VM (Rust 1.99, glibc 2.39, system Git
2.55), each tool ran five warmups and 60 measured invocations, rotating stock,
RELR and Git order with builds/tests idle.

These are full fork/exec/wait medians in milliseconds, measured with a small C
parent. The reports also retain daily-helper timings, standard deviations,
per-child peak RSS and page faults. Every fixture passes system `git fsck`.

| Profile / tiny command | Git ms | Stock grit ms | RELR probe ms |
| --- | ---: | ---: | ---: |
| release / branch | 1.237 | 2.177 | 2.144 |
| release / status | 1.524 | 2.699 | 2.662 |
| release / resolve | 1.127 | 2.292 | 2.264 |
| release / help | 1.006 | 3.102 | 3.064 |
| dist / branch | 1.182 | 2.013 | 1.858 |
| dist / status | 1.428 | 2.408 | 2.243 |
| dist / resolve | 1.068 | 2.083 | 1.926 |
| dist / help | 0.962 | 2.934 | 2.790 |

Tiny commands improve about **1–2% on release** and **5–8% on dist**. The
20,000-ref branch, 20,000-path status and 20,000-commit log cases stay close to
stock; the full report includes their variance and the differences in Git/grit
branch output limits. Release peak RSS falls by roughly 0.6–0.8 MiB. Dist RSS
does not consistently decrease: independent `/usr/bin/time -v` samples show tiny
branch/status/resolve median changes of +34/+58/+48 KiB, and help falls by 242 KiB.

RELR cuts relocation metadata from **793 KB to 14 KB** on release and **464 KB
to 9.9 KB** on dist. It fixes up the same 32,865 / 19,174 pointer slots on the
same 147 / 91 pages. Loader-only relocation cycles are mostly unchanged on dist;
binary layout shifts too, so the wall-time gain is not proof of a faster loader
relocation loop. No Unicode regex, filter, schema or tracing feature is removed.

The x86_64 GNU release now provides a separate `grit-x86_64-unknown-linux-gnu-relr.tar.gz`
asset for glibc 2.36+. The existing glibc 2.34 GNU artifact retains its build path
and remains the default and fallback. Stable and nightly installers always use
the existing GNU asset on glibc systems; RELR is a manual download only. The
final-tip review on `3fcfb24e9` found tiny-command regressions of about 1–2%
against a same-source stock build, confirmed by randomized and CPU-pinned
repeats. The earlier gains below do not establish performance of those artifacts.
Automatic selection is deferred until exact release artifacts demonstrate a
benefit. Both assets pass ELF hardening, startup and address randomization
checks; release CI runs them on 2.34 and 2.36 respectively. Other targets are
unchanged.

An earlier shipping-dist check on CLI/library source `aa1964221`
(4-vCPU Intel Xeon Platinum 8573C, 60 rotated warm runs per tool) measured:

| Tiny command | Git ms | Stock dist ms | RELR dist ms |
| --- | ---: | ---: | ---: |
| branch | 1.166 | 1.968 | 1.871 |
| status | 1.402 | 2.359 | 2.254 |
| resolve | 1.078 | 2.051 | 1.945 |
| help | 0.966 | 2.911 | 2.880 |

Branch/status/resolve improve **4.4–5.1%** here; help's 1.1% difference is within
the sample spread. Larger many-ref branch, wide status and deep log medians
improve 1.2–1.6%. Native peak RSS falls about 0.4–0.5 MiB in this layout;
the earlier dist memory caveat still applies. Loader-only cycles increase in
this run, so the wall-time result cannot be attributed solely to relocation
processing. The [shipping follow-up data](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/startup-relocations-shipping-dist.json)
retain all 20 scenarios, variance, RSS, faults, loader statistics and runtime
checks. Builds and tests were stopped throughout measurement.

A replacement-VM check rebuilt the exact reviewed `3fcfb24e9` CLI/library
source (installer correction `d8b6e83c3` changes no linked code). This host is
**AMD EPYC 7763**, 4 vCPU, Rust 1.99, glibc 2.39, system Git 2.55. With five
warmups and 60 rotated runs per tool and builds/tests idle, full-process dist
medians are:

| Tiny command | Git ms | Stock dist ms | RELR dist ms | RELR change |
| --- | ---: | ---: | ---: | ---: |
| branch | 1.412 | 2.138 | 2.177 | +1.83% |
| status | 1.713 | 2.529 | 2.581 | +2.07% |
| resolve | 1.280 | 2.188 | 2.209 | +0.97% |
| help | 1.148 | 3.172 | 3.201 | +0.88% |

All four paired-bootstrap 95% intervals for RELR-minus-stock median time are
above zero. Release differences are mixed and near zero (-0.20% to +0.40%).
Dist many-ref branch, wide status and deep log medians regress
0.54/0.33/0.39%, within their sample spread. Dist RSS is lower in this layout,
but startup is slower: smaller relocation metadata is insufficient grounds for
automatic selection. Fresh [dist data](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/startup-relocations-review-dist.json)
and [release data](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/startup-relocations-review-release.json)
include all 20 cases per profile, variance, peak RSS, faults, binary hashes,
ELF/ASLR checks, independent GNU-time samples and separate loader statistics.

The [complete report](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/startup-relocations.md)
links the methodology, ELF attribution and portability checks. Raw
[dist](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/startup-relocations-dist.json),
[release](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/startup-relocations-release.json),
[ELF metadata](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/startup-relocations-elf.json)
and [loader/RSS](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/startup-relocations-loader-rss.json)
reports retain the measurements. See [startup comparison methodology](https://grit-scm.com/docs/benchmarks/benchmarking.md/index.md#startup-relocation-comparison)
for reproduction commands.

## Ignore path normalization (2026-10-26)

Ignore queries used to canonicalize the same worktree root for every path. The library now checks the normalized path against the root already resolved when opening the repository, retaining the filesystem-alias fallback. On the daily 1,000-path fixture this removes **4,000 `readlink` calls** and **10,000 allocation calls**; heaptrack peak heap stays at **795,340 bytes**.

Baseline: `origin/main` at `c578daa58ced74ce4a33327942e29e8409b3cb10`; change: `786ec07eeb6b34f840112964d0dcf7f5d6ea4391`. Frozen before/after binaries and stock Git 2.55.0 were measured in one sitting on an Intel Xeon Platinum 8573C VM. Builds and tests were paused during timed measurements. Both the development release profile and the shipping `dist` profile were measured.

The daily harness uses its existing shell/xargs commands and 20 alternating samples per tool. Shipping-profile medians improve **8.84 → 7.11 ms** (Git: **2.50–2.53 ms**); helper-inclusive peak RSS is **14,624 → 14,560 KiB**. Release medians improve **9.27 → 7.29 ms** (Git: **2.50 ms**). The committed daily reports retain every sample.

The table below uses 50 direct process invocations per tool, rotating tool order after five warmups. These omit the daily shell/xargs overhead. RSS is the median of ten per-process peak readings from `/usr/bin/time -v`; all ten readings and wall-time samples are in `ignore-normalization-dist.json` and `ignore-normalization-release.json`. Git uses `check-ignore -q --stdin`; grit uses `ignore check --all`, with output discarded.

| Profile / workload | Git ms | Grit before ms | Grit after ms | Git peak KiB | Grit before peak KiB | Grit after peak KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| dist / `check-1000@synth-10k` | 2.21 | 7.56 | 5.72 | 4,644 | 10,312 | 10,218 |
| dist / `siblings-1000` | 1.25 | 5.39 | 3.99 | 4,506 | 9,974 | 10,004 |
| dist / `deep-1000` | 1.63 | 11.35 | 9.50 | 4,426 | 10,398 | 10,482 |
| dist / `tracked-1000` | 10.66 | 4.92 | 3.50 | 4,608 | 9,920 | 10,038 |
| dist / `check-1@synth-10k` | 1.02 | 1.55 | 1.53 | 4,618 | 9,500 | 9,484 |
| release / `check-1000@synth-10k` | 2.20 | 7.95 | 6.07 | 4,618 | 10,126 | 12,724 |
| release / `siblings-1000` | 1.25 | 5.80 | 4.29 | 4,460 | 9,776 | 12,532 |
| release / `deep-1000` | 1.61 | 11.82 | 9.96 | 4,444 | 10,382 | 13,038 |
| release / `tracked-1000` | 10.43 | 5.12 | 3.68 | 4,594 | 9,730 | 12,276 |
| release / `check-1@synth-10k` | 1.01 | 1.77 | 1.70 | 4,610 | 9,166 | 11,924 |

All five workloads improve in wall time. Shipping-profile RSS stays within the spread of repeated process measurements; the single-path case is near the timing noise floor. The non-LTO release build increases RSS by about **2.6 MiB** despite unchanged peak heap and lower allocation traffic. This profile-dependent binary footprint remains a limitation; this change does not claim to close the memory gap to Git.

The daily fixture queries 1,000 distinct parents, mostly absent, in the standard 10k-file repository with an emptied index and `*.txt`. Auxiliary fixtures use 1,000 existing files across ten directories, 1,000 files below twelve nested directories, the same wide fixture staged with `git add -f .`, and a single path from the daily fixture. All auxiliary ignore files contain `*.txt`.

Reproduce the target with each frozen binary:

```bash
cargo build --release -p grit-cli -p grit-utils
cargo build --profile dist -p grit-cli -p grit-utils
grit-bench daily --only ignore/check-1000@synth-10k --grit /path/to/frozen/grit --output ignore.json
```

## Ignore exclude stacks per directory (2026-10-31)

`grit ignore check` (and every other `IgnoreMatcher` caller) used to rebuild its view of the per-directory `.gitignore` stack for each path. It cloned the ancestor directory list, assembled a fresh vector of every pattern group, and re-probed every ancestor directory against the patterns to see whether it was excluded. When a path's parent was new, it also stat'd every ancestor to make sure none was a file. Parsed `.gitignore` files were already cached, so no file was read twice; the cost was all per-path bookkeeping and system calls.

The matcher now keeps one state per directory, derived from its parent's state the first time a path inside it is queried. That state records the exclude files that apply and whether the directory or an ancestor is already excluded, so each directory is stat'd, read and probed once. Each pattern group is scanned from its last line, and the base directory and basename are stripped once per group, without allocating. The batch check reuses the `stat` it already makes for each queried path to learn that no ancestor is a regular file, so it makes one `stat` per path, as Git does. `RepoPathNormalizer` resolves the working directory's position once and normalizes each relative path lexically with a single allocation. Results are byte-identical in human, `--json` and `--markdown` output.

Release builds, Linux VM with four logical CPUs, `git version 2.43.0`, measured in one sitting with frozen before and after binaries:

| Workload | Git ms | Grit before ms | Grit after ms | Git peak KiB | Grit before peak KiB | Grit after peak KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| daily `ignore/check-1000@synth-10k` | 2.25 | 6.99 | 5.56 | 6,592 | 18,064 | 18,020 |
| nested `.gitignore` per directory, 1,000 paths | 49.2 | 7.5 | 5.3 | 6,772 | 15,756 | 15,108 |
| nested `.gitignore` per directory, 10,000 paths | 920 | 40.6 | 18.8 | 6,800 | 20,564 | 18,492 |
| `grit-bench ignore` (`ignore-check-1000`) | 1.9 | 4.4 | 3.5 | | | |

The daily row runs through the harness's `xargs` wrapper, so it includes that overhead. The nested rows use synth-10k with a root `.gitignore` of ten patterns (negation, `**/cache/**`, a character class), a five-line `.gitignore` in each of its 100 directories, and `info/exclude`. The 1,000 queries cover tracked files, plain and directory-only matches, ignored ancestors, re-included names, and missing subdirectories; the 10,000-path row repeats that list ten times. Those rows are `hyperfine -N` means for `ignore check --all` and `git check-ignore -v -n`, and median wall time for the 10,000-path row, where the argument list is too long for `hyperfine`. On the daily fixture, `stat` calls drop from 2,077 to 1,077 (Git: 1,014) with the same 1,015 opens. On the 1,000-path nested fixture, `stat` calls drop from 1,376 to 1,166, `malloc` calls from 60,111 to 24,428, and instructions (callgrind) from 67.7M to 41.7M.

Peak RSS barely moves because almost none of it is heap. A one-path query in a one-file repository peaks at about 13 MiB, of which less than 1 MiB is heap. The rest is pages of the 45 MB binary touched while loading and relocating it, which also accounts for most of the remaining wall-time gap to Git: `grit --version` alone takes about 1.5 ms. The committed [before](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-ignore-dir-stack-release-before.json) and [after](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-ignore-dir-stack-release-after.json) daily reports keep every sample. Both record the checkout's commit as `grit_commit`, not the frozen binary's.

```bash
cargo build --release -p grit-cli -p grit-utils
grit-bench daily --only ignore/check-1000@synth-10k --grit /path/to/frozen/grit --min-runs 40 --format json
grit-bench ignore --grit /path/to/frozen/grit --min-runs 30 --format json
```

## Log statistics on packed history

`grit log --limit 50 --stat` reuses its batched file counts for the JSON summary,
without computing a second full patch. The library trims identical boundary lines
before Myers counting, interns borrowed LF lines into compact ids, transfers owned
blob buffers into diff inputs, and borrows valid UTF-8 while counting. CR-containing
inputs retain the existing tokenizer and lossy decoding remains unchanged.

Factory VM, 2026-10-26: four logical CPUs, AMD EPYC 7763, 16 GiB RAM, Linux x86_64,
Rust 1.99 `release` profile, stock Git 2.55.0. The daily fixture is upstream git.git
at v2.47.0 (`777489f9e09c8d0dd6b12f9d90de6376330577a2`). Frozen binaries compare
origin/main `f2a97fc88fe4` with `f3c83c08a945` in the same sitting. Builds and tests
are idle during measurement. Hyperfine uses two warmups and ten measured runs per
tool; RSS is the maximum of three separate `/usr/bin/time -v` runs.

| Command | Mean wall time | Peak RSS |
| --- | ---: | ---: |
| Stock `git log --stat -50` | 103.1 ms | 35.5 MiB |
| Grit before | 385.2 ms | 112.5 MiB |
| Grit after | 192.5 ms | 97.6 MiB |

That is about **50% less time and 13% less peak RSS**. Separate Heaptrack runs count
**348,505 → 80,281 allocations**, with peak live heap **53.77 → 49.61 MB** (decimal).
The [before](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/log-stat-before.json)
and [after](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/log-stat-after.json)
daily reports cover all eleven log scenarios and eight related diff, show and
shortlog scenarios. Those reports use medians and helper-inclusive peak RSS, so
their values differ from the direct process measurements above. Raw direct samples,
RSS samples and Criterion estimates are in
[the measurement record](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/log-stat-study.json).

Thirty randomized interleaved samples per tool checked startup-heavy commands:
JSON log **6.04 → 5.59 ms**, oneline log **5.96 → 5.63 ms**, author-filtered log
**6.54 → 6.01 ms**, and show **4.06 → 3.39 ms** (medians). Direct RSS checks still
show about **1 MiB more** for oneline/JSON log. The daily table records similarly
small memory increases on other short commands. On 2026-10-27 the maintainer
accepted this deliberate, documented tradeoff for the target's time and memory
improvements. Removing extra startup pages from generic diff instantiations is a
separate follow-up: candidates include a non-generic inner function, dynamic
dispatch for the counting sink, or keeping instantiations out of startup paths.
This follow-up is separate from relocation-cost work in #4626.

A review recheck on 2026-10-27 rebuilt both original source snapshots on a fresh
four-logical-CPU AMD EPYC 9V45 VM. With 100 randomized interleaved samples per tool,
JSON log was **3.650 → 3.625 ms** and author-filtered log with limit 50 was
**2.819 → 2.806 ms** without a commit graph and **2.914 → 2.901 ms** with one.
The apparent short-command slowdowns in the adaptive table did not persist.
A separate 20-sample interleaved target check was **263.9 → 133.8 ms** versus
Git **75.1 ms**, with direct peak RSS **117.4 → 98.8 MiB** (maximum of five runs).
The [review recheck](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/log-stat-review-recheck.json)
retains the raw samples and exact commands. These comparisons each use one VM
and one sitting; their absolute times are not interchangeable across machines.

Criterion compares the unchanged full-file counter with `line_change_counts` in
the same process (20 samples): **19.74 → 3.17 ms** for a small edit surrounded by
100,000 unchanged lines, and **139 → 78 µs** for 1,000 entirely replaced lines.
The latter checks the case where boundary trimming cannot help.

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench --grit ./target/release/grit daily \
  --only log/stat-n50@git-git --output /tmp/log-stat.json
cargo bench -p grit-lib --features http-ureq,tracing --bench log_stat -- \
  --sample-size 20 --warm-up-time 1 --measurement-time 2
```

## Clean wide content merges

Issue [#2036](https://github.com/gitbutlerapp/grit/issues/2036) measures 20,000 small text files
where one side prepends a line and the other appends one. Rename detection and result-tree
composition now borrow existing paths and entries, parsed trees are shared with the ODB,
and checkout write jobs borrow their diff records. Side maps are released before composing
the result tree.

Factory VM, 2026-10-26: 4 vCPU AMD EPYC 7763, 16 GiB RAM, Linux x86_64,
Rust 1.99 `release` profile, system Git 2.55.0. One warmup precedes four interleaved
fresh-process runs of Git, origin/main (`74adad7a`), and the fix in the same worktree,
reset before each merge. Compilation is paused. Wall time uses Python `perf_counter`;
peak RSS uses `/usr/bin/time -v`. The packed row reproduces the existing daily
`merge/content-both-sides@wide-20k` fixture (200 directories, 100 files each, `git gc`
before merging). The loose rows follow the issue's reproduction at 10k and 20k files.

| Files / layout | System Git | Grit origin/main | Grit fixed |
| --- | ---: | ---: | ---: |
| 10k loose | 1,923 ms / 19.4 MiB | 1,525 ms / 55.8 MiB | 1,482 ms / 40.9 MiB |
| 20k loose | 3,793 ms / 33.9 MiB | 2,986 ms / 96.0 MiB | 2,940 ms / 61.8 MiB |
| 20k packed (daily) | 2,471 ms / 48.1 MiB | 2,832 ms / 107.5 MiB | 2,700 ms / 74.7 MiB |

Times are medians; RSS is the maximum of the four measured runs. Every result tree
matches system Git, the index and worktree remain clean, and `git fsck --strict` passes.
Raw samples and both baselines are committed in
[`merge-content-memory.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/merge-content-memory.json).

Separate Heaptrack profiles put peak live allocations at **57.20 → 37.54 MB** for
loose objects and **67.22 → 47.63 MB** for the packed fixture (decimal MB). The
original peak retained merge maps; after removing those copies, checkout's repeated
diff records became material and were changed to references. The RSS table excludes
profiler overhead.

Criterion's `merge_content` bench exercises the library tree merge without checkout:
**1,345.8 ms before → 1,292.8 ms after** (10 samples, about 4% faster). Run it with:

```bash
cargo bench -p grit-lib --bench merge_content
```

Linux regression tests measure tree merging and full integration in separate fresh
processes, including loose and packed objects. The measured debug API peaks were
about **56/58 MiB** for loose tree/full merges and **74/82 MiB** for packed tree/full
merges. Their budgets are **65 MiB loose** and **88 MiB packed**, allowing allocator
variation while rejecting the old loose merge's roughly 78/95 MiB peaks. Packed RSS
includes mapped pack pages and bounded object/delta caches. Git still uses less memory;
the regression budgets preserve the measured improvement.

## Local unshallow with historical tags

Issue [#2053](https://github.com/gitbutlerapp/grit/issues/2053) covers tags entering newly fetched history. Local fetch now selects following tags using the shallow response's effective boundaries. Lightweight targets arrive through the branch roots; missing annotated tag objects and chains get additional wants. Already-local tags still skip the reachability walk.

Factory VM, 2026-10-26: 4 vCPU AMD EPYC 9V74, 16 GiB RAM, Ubuntu 24.04, Rust 1.99 `release`, system Git 2.55.0. The packed `file://` fixture has 3,000 linear commits, 1,500 tags on even-numbered commits and 1,000 extra tip tags. Each sample starts from a fresh system-Git depth-1 clone outside the timed region. Complete no-op clients are first unshallowed by Git. One warmup precedes seven interleaved runs per binary, with rotating order and builds/tests paused. Wall time uses Python `perf_counter`; peak RSS uses `/usr/bin/time -v`. Automatic maintenance is disabled for all binaries.

| Operation | System Git | Grit main | Grit fixed |
| --- | ---: | ---: | ---: |
| One unshallow fetch | 226.7 ms / 8.6 MiB | 213.3 ms / 23.7 MiB | 492.0 ms / 25.8 MiB |
| Complete history and tags | 226.8 ms / 8.6 MiB | 537.6 ms / 23.3 MiB | 495.9 ms / 26.1 MiB |
| Complete no-op fetch | 51.8 ms / 7.0 MiB | 26.5 ms / 13.8 MiB | 25.9 ms / 15.3 MiB |
| Shallow no-op fetch | 26.9 ms / 6.1 MiB | 32.8 ms / 14.1 MiB | 32.3 ms / 15.6 MiB |

Times are medians; RSS is the largest peak across seven runs. Main is `e10563aae0fde79b96e13cca50d170d3d35e7b7b`. **Main's single unshallow result is incomplete:** it omits 1,499 historical tags. The complete-result row includes its required second fetch, including both commands' elapsed time and the larger process RSS. Git and fixed Grit produce all tags in one operation. No-op timing remains within the observed sample spread; following tags does not add a history scan when every tag is already local. RSS includes the tag updates retained during object copying in the fixed single-fetch operation. In ten separate complete no-op runs sampled through `/proc/<pid>/status`, median anonymous RSS stayed about 1.8 MiB before and after; the measured no-op RSS increase is in file-backed pages. All final results pass `git fsck --strict`.

Raw timing and RSS samples, fixture details and commands are recorded in [`fetch-local-unshallow-tags.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/fetch-local-unshallow-tags.json).

## Local fetch with many refs

Local fetch reuses the repository's ref store for tag checks, tracking-ref classification, and object-copy exclusion tips. Exclusion tips are collected only when the finalized shallow/deepen request needs an object copy. Up-to-date fetches still perform pruning and fetch metadata maintenance; loose refs override packed refs, symbolic refs resolve through the store, and both files and reftable backends retain their freshness checks.

Factory VM, 2026-10-26, 4 vCPU, Rust 1.99, `release` profile, system Git 2.55.0. The issue #2037 fixture contains one commit, 3,000 extra branches and 2,000 tags, with packed remote refs and a system-Git-cloned client. The deep-history fixture uses upstream `git/git` at `7e34d8174dcda3435d1404c9a30bef46afe6c13e` (82,563 commits), its original refs, and the same extra branches and tags. Every command fetches from a local bare repository without transferring objects or changing refs.

| Fixture / command | Git mean ms | Grit main mean ms | Grit after mean ms | Git peak KiB | Grit main peak KiB | Grit after peak KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| One commit, `fetch origin` | 41.3 | 219.1 | 35.5 | 9,024 | 17,744 | 16,456 |
| One commit, `fetch --no-tags origin` | 37.2 | 210.6 | 24.1 | 8,684 | 17,188 | 16,268 |
| Deep history, `fetch origin` | 47.9 | 219.6 | 41.0 | 43,844 | 19,696 | 19,080 |
| Deep history, `fetch --no-tags origin` | 40.7 | 203.7 | 24.7 | 43,664 | 19,504 | 19,016 |

The baseline is `origin/main` at `a61c6cf784d65e54c9f8105b0d8f66e0403dccee`. Frozen binaries for that revision and the working-tree fix were measured alongside stock Git in one session. Run `hyperfine -N --warmup 3 --runs 10` with those three binaries, first with `fetch origin`, then with `fetch --no-tags origin`, from each client. Builds and tests were paused for measurements. Peak RSS comes from separate `/usr/bin/time -v` runs; the table uses the larger of the two Git RSS samples. Individual wall-time samples, spread, and instrumentation are committed in `grit-utils/baselines/fetch-local-many-refs.json`.

Separate `strace -f -e trace=openat` runs counted all open attempts, including missing files, for paths ending in `/config` or `/packed-refs`. On the one-commit fixture, default Grit fetch went from **9,039 config opens and 3,008 packed-refs opens to 34 and 6**. Git used 35 and 4. With `--no-tags`, Grit went from 9,037 and 3,007 to 34 and 6. Cache validation and loose-ref lookup remain enabled.

The daily scenarios `fetch/noop-local@many-refs` and `fetch/noop-local-no-tags@many-refs` reproduce this shape with 20 alternating samples per tool, check unchanged refs and `git fsck --strict`, and require no cached network fixture. Their committed before/after daily reports include helper-inclusive RSS, which differs from the direct process measurements above. Reproduce them with:

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench --grit ./target/release/grit daily \
  --only fetch/noop-local@many-refs,fetch/noop-local-no-tags@many-refs \
  --output /tmp/fetch-local-many-refs.json
cargo bench -p grit-lib --bench fetch_local_many_refs
```

Criterion mean library times (20 samples, one-second warmup, three-second measurement target):

| Tag mode | Before ms | After ms |
| --- | ---: | ---: |
| Following | 215.43 | 31.64 |
| None | 203.20 | 20.73 |
| All | 354.65 | 47.79 |

The Criterion suite exercises the public local fetch (`repo.remotes().fetch_from(TransferEndpoint::Local(..))`, reopening the repository each iteration) with `Following`, `None`, and `All` tags on a system-Git-created 5,000-ref fixture. Compatibility tests also cover changed remote tips, loose overrides, symbolic refs, pruning followed by a no-op fetch, linked worktrees, reftable, and explicit `FETCH_HEAD` maintenance. CI checks behavior rather than elapsed-time thresholds.

## Clone and fetch of 200,000 branches (2026-10-28, #2189)

Writing many new refs used to be quadratic. The directory/file name check compared every pair of names in a batch. A clone wrote each imported ref as a loose file, with its own lock, then packed most of them in a second pass, and a bare clone's `refs/heads/*` were never packed. Now the batch check probes a sorted set, so it costs `O(n log n)`. A clone commits its imported refs as an initial transaction ([`RefTransaction::initial`](https://docs.rs/grit-lib/latest/grit_lib/refs/store/struct.RefTransaction.html#method.initial)): one `packed-refs` write, the same layout Git produces. An ordinary fetch still writes loose refs, as Git does, but makes far fewer system calls per ref. It takes each ref's lock once, revalidates `packed-refs` once per check, looks up shared ancestors once per batch and probes a missing ref with a single `lstat`. It also passes the batch to the store without copying it.

The fixture is the issue's bare repository: one commit and N branches (`refs/heads/b/n<i>` plus `main`) in `packed-refs`. Each run starts from a fresh destination over `file://`. Fetch goes into an empty bare repository with `+refs/heads/*:refs/remotes/origin/*`. Measured on 2026-10-28: 4 vCPU Intel Xeon, 16 GiB RAM, ext4, Rust 1.99, `release` profile, system Git 2.43.0. Main is `dc9208904`. Git, main and the fix alternate within each round (five rounds at 25k, three at 200k). Medians come from `/usr/bin/time`, and peak RSS is the largest across runs. Every result has the full ref count and passes `git fsck --strict`.

| Command / branches | Git ms / KiB | Grit main ms / KiB | Grit after ms / KiB |
| --- | ---: | ---: | ---: |
| `clone --bare`, 25,000 | 190 / 31,260 | 2,840 / 63,228 | 150 / 46,164 |
| `fetch` into a bare repo, 25,000 | 1,120 / 28,104 | 2,720 / 68,240 | 1,480 / 56,516 |
| `clone --bare`, 200,000 | 1,570 / 208,616 | 43,630 / 396,552 | 1,420 / 260,408 |
| `fetch` into a bare repo, 200,000 | 10,450 / 184,572 | 42,490 / 432,436 | 12,430 / 337,260 |

Fetch user time is now linear in the ref count: about 0.4 s at 25,000 refs and 3 s at 200,000, where main took 0.9 s and 25.6 s. What remains of the fetch gap is creating 200,000 loose files in one directory. Git pays that too: about 80% of its time is system time. Per ref, grit now makes 20 system calls, down from 62 on main. Git makes 42. Raw samples are in [`clone-fetch-many-branches-2189.json`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/clone-fetch-many-branches-2189.json). To reproduce, build the fixture with the commands in the issue, then time `grit clone --bare file://$PWD/tiny g1` against `git clone -q --bare`, and `grit -C f1 fetch origin` against `git -C f2 fetch -q origin`.

## Bounded branch containment on 10,000 unique tips

Issue #2035 is measured on a full clone of git.git at `7e34d8174dcda3435d1404c9a30bef46afe6c13e` (82,563 reachable commits). Create 10,000 distinct `refs/heads/dogfood/NNNNN` tips from `git rev-list -n 10000 HEAD`, then run `git pack-refs --all`. This fixture has no commit-graph. The review correction was measured on 2026-10-26 against stock Git 2.55.0: Ubuntu 24.04.5, four vCPUs (AMD EPYC 7763), 16 GiB RAM, Rust 1.99.0, `release` profile. Main is `a61c6cf784d65`, reviewed is `9d66d522369c`, and corrected is `da77132c09dd`. No compiler or test suite runs during measurement.

```bash
grit branch --contains HEAD~100 -n 100 --json
git for-each-ref --contains=HEAD~100 --count=100 \
  '--format=%(refname) %(subject) %(committerdate:iso8601) %(upstream)' refs/heads/
# Human comparison: omit --json and use --format='%(refname:short)' for Git.
# Unlimited comparison: omit -n / --count; unfiltered control: omit --contains too.
```

A 100-row JSON query warms every executable before alternating measurements. Limits 1 and 100 and the unfiltered control use three alternating Git/main/reviewed/corrected samples; limit 1,000 and unlimited containment use one sample because Git scans the whole ref inventory. Wall time surrounds the subprocess, and peak RSS comes from `/usr/bin/time -v`. Ordered JSON and human refs and JSON truncation are checked against system Git. The before/reviewed/after baseline JSON files retain every timing sample, maximum RSS, and the measured commit in `tools.grit_commit`.

| Output / limit | Git ms / KiB | Main ms / KiB | Reviewed ms / KiB | Corrected ms / KiB |
| --- | ---: | ---: | ---: | ---: |
| JSON / 1 | 33.82 / 19,164 | 2608.18 / 100,924 | 14.74 / 26,996 | 14.30 / 27,120 |
| JSON / 100 | 199.07 / 31,312 | 2612.84 / 101,288 | 34.28 / 34,412 | 33.91 / 34,668 |
| JSON / 1000 | 34184.82 / 70,880 | 2630.10 / 100,904 | 249.73 / 59,900 | 232.96 / 60,352 |
| JSON / unlimited | 34139.51 / 70,872 | 2622.00 / 101,032 | 2366.62 / 102,608 | 2389.44 / 102,420 |
| JSON / unfiltered, unlimited | 158.39 / 25,604 | 182.69 / 37,100 | 180.51 / 38,560 | 180.89 / 38,644 |
| Human / 1 | 34.27 / 19,136 | 2637.82 / 101,036 | 14.82 / 26,992 | 14.73 / 26,932 |
| Human / 100 | 198.35 / 31,152 | 2608.95 / 102,092 | 33.59 / 34,020 | 33.54 / 34,288 |
| Human / 1000 | 34608.55 / 70,668 | 2636.97 / 102,468 | 227.79 / 59,976 | 229.86 / 60,240 |
| Human / unlimited | 34405.21 / 70,732 | 2625.78 / 102,072 | 2393.66 / 102,100 | 2352.27 / 102,412 |
| Human / unfiltered, unlimited | 161.70 / 5,240 | 135.36 / 19,816 | 134.94 / 19,688 | 134.11 / 19,720 |

The finite name-ordered path shares positive and negative ancestry between evaluated tips and stops after enough matches, including the CLI's extra truncation probe. Without a commit-graph, expensive negative probes can meet an incrementally explored boundary of proven target ancestors. Committer dates order that exploration but never exclude ancestry. Unlimited queries, version/date sorts and combined filters retain the batch path. Grafts and enabled replacement refs also use the batch path: rewritten parents can form cycles, so a target's ancestors are not necessarily nonmatches. Commit-graph linkage and generations are ignored for these rewritten histories.

The correction retains the reviewed build's bounded-query speed and memory use on ordinary histories; the unlimited and unfiltered controls remain comparable. Rewritten histories use the general batch path for correctness.

The original bounded-query implementation (`9d66d522369c`) measured these means on its synthetic 10,000-tip Criterion fixture (10 samples, 0.1-second warmup and 0.2-second requested measurement, extended automatically for long queries):

| Matching-row limit | No commit-graph ms | Commit-graph ms |
| --- | ---: | ---: |
| 1 | 6.02 | 6.18 |
| 100 | 7.10 | 6.86 |
| 1,000 | 64.78 | 31.16 |
| Unlimited | 284.22 | 184.35 |

Run the library fixture with `cargo bench -p grit-lib --bench ref_query_contains_tags`. Its `ref_query_contains_10k_unique_branches` group covers limits 1, 100, 1,000 and unlimited with and without a commit-graph, opening a fresh repository handle per sample. The daily command matrix includes `branch/contains@git-git-10k-refs` and `branch/contains-json@git-git-10k-refs`, using an isolated copy of the pinned git.git fixture.


## Linux worktree metadata scans (2026-10-26)

Status and worktree diffs used to construct an absolute path for each indexed file and repeat the lookup of its parent directories. On 64-bit Linux, larger scans now open each eligible parent once with `O_PATH` and inspect its file names with `fstatat` and `AT_SYMLINK_NOFOLLOW`. The untracked walk also reuses the operation's configuration instead of loading the cascade again to choose its worker settings. Index stat fields and file-type handling stay the same. Scans below 1,000 entries, directories below eight files, and case or Unicode path resolution keep the existing metadata lookup.

The [before](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/status-scan-before.json) and [after](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/status-scan-after.json) reports compare release builds based on `a61c6cf` with system Git 2.55.0 on a four-CPU AMD EPYC VM using ext4. Initial reproduction used `grit-bench daily --only status/clean@synth-10k` with 20 runs per tool. The final comparisons reuse the daily fixtures with warm indexes and rotate Git/before/after through `grit-bench measure-direct-run`: 30 runs per tool, 200 for small status, 10 for deep-tree, and 60 for mutating canaries. Both Grit binaries have the same basename and equal-length parent paths (`/tmp/grit-4658-binaries/{before,after_}/grit`); copied binaries were verified byte-for-byte against the builds. Times are medians; RSS is the maximum across the runs, cross-checked with ten `/usr/bin/time -v` runs per tool.

| Scenario | Git ms | Grit before ms | Grit after ms | RSS before KiB | RSS after KiB |
| --- | ---: | ---: | ---: | ---: | ---: |
| `status/clean@synth-10k` | 13.321 | 13.564 | 11.684 | 19744 | 19564 |
| `status/clean@small` | 1.670 | 2.831 | 2.778 | 15656 | 15480 |
| `status/clean@synth-100k` | 116.114 | 104.294 | 84.944 | 36104 | 35896 |
| `status/clean@deep-tree` | 2062.934 | 1024.723 | 1014.167 | 75236 | 75404 |
| `diff/worktree@synth-10k` | 11.841 | 10.873 | 9.061 | 19768 | 19496 |

Heaptrack on the 10k-file fixture counted **70,813 → 50,252 allocations** (29% fewer), and **31,001 → 11,662 temporary allocations** (62% fewer). Peak live heap stayed about 2.08 MB; process RSS remains broadly unchanged. The full reports include every related status case and the diff/staging canaries.

Related `diff`, `add`, and `commit` rows are included in the reports. The commit canary compares auto-staging `grit commit -m bench` with `git commit -am bench`; preparation and resets stay outside the timed region. The current daily `status/renamed` preparation resets to a clean tree, and `status/paths` selects `data/d0000`, absent from the current `synth-100k` fixture; those registry cases are reported as they currently run. Rust integration tests separately cover actual renames and raw-byte paths.

## Shared environment snapshots

Repository discovery and config loading clone an explicit `Environment`. Its captured
variables share immutable storage, while typed overrides remain independent. This removes
repeated copies of inherited variables from short commands.

The isolated check on 2026-10-27 compares original base `f23547129` with implementation
`b04e11d54`, both built with Rust 1.99.0 and `cargo build --release -p grit-cli -p grit-utils`
in **the same source and target directory**, forcing recompilation between archives and
checking executable hashes. The VM has four vCPUs on AMD EPYC 9V45, Linux, and system
Git 2.55.0. The daily `many-refs` fixture has HEAD `b2c0606217`, 10,000 generated refs
and one local branch in the working copy.

| Operation | Git | Grit before | Grit after |
| --- | ---: | ---: | ---: |
| Branch JSON, 600 rotating direct-spawn samples, median | 1.468 ms | 2.153 ms | 2.074 ms |
| Branch JSON, hyperfine in both orders, pooled median | 1.382 ms | 2.108 ms | 2.000 ms |
| Branch creation, 120 rotating direct-spawn samples, median | 1.371 ms | 2.353 ms | 2.246 ms |
| Branch JSON peak RSS, median of 20 `time -v` runs | 5,738 KiB | 14,318 KiB | 14,042 KiB |

A fresh baseline reconstruction produced the identical executable hash. A final 600-round
check measured 2.158 ms for the preserved base, 2.126 ms for its reconstruction, and
2.075 ms for the patch (a 2.4–3.8% gain); Git measured 1.477 ms.

The Git headline comparison is `git branch -vv`, matching the daily scenario. The repeated
daily matrix uses 40 samples per revision in opposite revision orders, including matching
`grit-bench` library drivers for rename and merged-ref. None of its timing increases exceeds
5%; all eight branch scenarios improve in the separate 120-sample direct comparisons.
Heaptrack allocation calls for branch JSON fall from 5,582 to 4,477; peak live heap stays
about 325 KB, set earlier by argument parsing. Human, JSON and Markdown output is unchanged.

This is **not a zero-regression result**. In 20 alternating RSS samples, tag listing rises
12,596 → 12,934 KiB (+2.7%), config get 10,820 → 10,934 KiB (+1.1%), clean status
19,152 → 19,210 KiB (+0.3%), and the merged-ref helper 10,408 → 10,504 KiB (+0.9%).
Tag peak live heap falls about 1.24 → 1.22 MB, with 22,064 → 20,962 allocations. Earlier
VM checks observed tag +2.1–2.8%, status up to +0.9% and config up to +1.9%; the maintainer
approved review with these trade-offs disclosed. Allocator and startup residency work is
tracked under [#4626](https://github.com/gitbutlerapp/grit/issues/4626).

The earlier end-to-end comparisons used newer main builds and do not isolate this patch.
A reviewer measured a roughly 7% slowdown against the original base on EPYC 7763; the
same-path experiment above does not reproduce it on this VM. These results do not establish
whether CPU or build layout explains that discrepancy. Reproduce the isolated comparison
with `bash scripts/bench-branch-snapshots.sh BEFORE AFTER FIXTURE NEW_OUTPUT_DIRECTORY`;
archive timestamps are refreshed explicitly to prevent stale Cargo artifacts.
The full measurements, hashes and limitations are in
`grit-utils/baselines/branch-json-4643-isolated.json` and the measurement ledger. The
committed daily before/after baselines now contain the isolated 40-sample results.
For the library clone benchmark, run `cargo bench -p grit-lib --bench environment`.

## grit-bench fixture cache

Scenario suites reuse prepared repositories under **`GRIT_BENCH_CACHE`** (default `$XDG_CACHE_HOME/grit-bench`). **`grit-bench fixtures prepare`** clones upstream **git.git** at tag **`v2.47.0`** (bare, verified `HEAD` oid) and generates synthetic tiers **S**, **M**, and **L** with the grit-lib generator. **`grit-bench fixtures list`** prints identity (name, tier, `HEAD`, object/ref/file counts). Mutating runs take a **`git clone --local`** working copy so the cache stays read-only. JSON reports may include a top-level **`fixtures`** array with the same identity fields.

```bash
./target/release/grit-bench fixtures prepare --format json
./target/release/grit-bench fixtures list --format markdown
```

### Daily matrix: partial staging on `large.txt`

Two **`grit-bench daily`** worktree scenarios share the same prepare hook (`prepare-worktree-add-hunks`): reset index/worktree to **`HEAD`**, then apply **500** one-line edits on a committed **50 000**-line `large.txt`. Each timed iteration runs that prepare again so partial staging from the previous sample does not carry over.

| Scenario id | Grit (timed) | Git (timed) |
| --- | --- | --- |
| **`add/hunks@large-file`** | `grit add --hunks` with the first **half** of unstaged hunk ids (same selection grit uses when writing bench artifacts) | `git diff -- large.txt`, keep the first **half** of unified hunks in file order, pipe to **`git apply --cached -`** |
| **`add/patch-file@large-file`** | `grit add --patch-file` on the prebuilt half-hunk unified diff written during prepare | **`git apply --cached`** on the same prebuilt patch (apply-only) |

Acceptance for both: grit median wall time and peak RSS must be **≤** git on the same scenario.

### Path-limited history

On 2026-10-26, release builds at upstream `818de8006` and the path-walk optimization were compared against system Git 2.55.0 on a four-vCPU AMD EPYC 7763 VM. The pinned git.git fixture is `v2.47.0`. Each tool had one warmup, then 10 alternating samples (30 for short commands) through the same `grit-bench measure-direct-run` helper used by the daily harness. The table reports median wall time and maximum peak RSS; GNU `time -v` independently checked the target.

| Scenario | Git ms | Grit before ms | Grit after ms | RSS before KiB | RSS after KiB |
| --- | ---: | ---: | ---: | ---: | ---: |
| `log/path@git-git` | 370.22 | 3229.01 | 462.62 | 312,212 | 273,720 |
| `log/limit-path@git-git` | 25.01 | 62.81 | 20.58 | 40,668 | 36,280 |
| `revs/count-path-head@git-git` | 453.37 | 2512.87 | 513.56 | 311,044 | 272,028 |

Path-limited history now scans the requested entries directly in reusable tree buffers, retains packed delta bases, and stops comparing merge parents once the selected path is unchanged. Commit-object caching remains available to author filters. The [full results](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/baselines/daily-log-path-4526.json) include all eleven log scenarios and related operations, with individual samples and the original build recorded as `baseline_grit`. The `revs/count-path-head` row uses pinned `HEAD`; the daily row’s historical revision `7e34d81` is absent from this fixture.

## Branch inventory and packed refs

Branch inventory shares one validated configuration snapshot across upstream lookups and reuses the repository's ref store for the main worktree's `HEAD`. Full packed-ref scans collect records in file order before building the sorted cache, avoiding an intermediate hash table. Ranged scans validate the records they visit and both neighbours, falling back to a full scan when they encounter an ordering violation. The [refs guide](https://grit-scm.com/docs/benchmarks/library/refs.md/index.md#reading-packed-refs) describes the approved trust model.

The October 27 comparison freezes main at `aa1964221` and applies only the branch-inventory changes. It uses Rust 1.99.0's unchanged `release` profile, stock **Git 2.55.0**, and a four-vCPU **AMD EPYC 9V74** VM. The binaries remain PIE executables. Each row has **60 shuffled, interleaved runs per tool**, ten warmups, matched executable basenames and path lengths, and no concurrent builds or tests. Peak RSS is the maximum of ten `/usr/bin/time -v` runs. Preparation stays outside timed regions.

The authoritative samples are `branch-inventory-4184-before.json` and `branch-inventory-4184-after.json`. Separate `daily-branch-inventory-4184-*.json` files exercise the registered daily scenarios with `grit-bench`; those sequential report captures are supplementary to the interleaved comparison.

| Operation | Git ms | Grit before ms | Grit after ms | RSS before KiB | RSS after KiB |
| --- | ---: | ---: | ---: | ---: | ---: |
| Daily branch listing | 1.402 | 2.168 | 2.120 | 12,604 | 12,756 |
| List 6,667 packed local branches | 22.574 | 69.599 | 11.343 | 16,432 | 16,312 |
| List 99,999 local branches in a 100k-ref repo | 351.659 | 949.174 | 151.917 | 83,040 | 82,952 |
| List all 100k refs through the library driver | 58.567 | 152.284 | 127.940 | 54,548 | 51,028 |
| List 6,667 loose local branches | 86.359 | 199.979 | 140.001 | 16,196 | 16,092 |
| Clean status, small repo | 1.902 | 2.809 | 2.813 | 15,252 | 15,340 |

The daily clone has one local branch and roughly 10k remote refs and tags. The local-branch rows use the fixture sources themselves; the loose source really retains loose files. The two 100k-ref repositories retain the source's objects and history, with sorted refs pointing at `HEAD`: one has 66,666 remote refs, 33,333 tags and `main`; the other has 99,999 local refs and one remote ref. System `git fsck --strict --no-dangling` checks both.

The profile explains the large-repository gains: listing 6,667 local branches reduces **33,410 → 61 `statx` calls**, **419,065 → 118,849 allocations**, and **44.0 → 11.7 MB of requested allocation bytes**. Small status also does less work: **6,930 → 6,621 allocations** and **526 → 515 syscalls**. The file-order collector reduces requested allocation bytes for the 100k full scan from **86.8 to 80.1 MB**; peak RSS falls about **6%** there.

Short-command differences need caution. `branch-inventory-4184-bisection.json` records independent config-reuse, file-order and hybrid-validation builds, plus executable-file controls. Reordering two helpers produced a **byte-identical** baseline executable, yet separate files differed by about **9%** on status; a hard link stayed within **1%**. Copying the current baseline changed status by about **4%**.

The hybrid and file-order pieces add no allocations or syscalls to small status, and the combined current-main status result is flat. The small positive deltas in create and tag rows include zero in paired bootstrap 95% intervals. Treat the headline's 2% lower median as near-flat; the large-repository gains are much larger than these controls.

A separate merged-ref control changed by about 6% when copying an identical baseline, while its hard link stayed flat; the after binary was about 4% slower in that control capture. Short-row RSS also remains sensitive to executable mapping: the same baseline varied by about 0.4 MiB between captures. All values are retained in the evidence, including the small RSS increases on short rows. Startup/file-placement investigation belongs to #2028; this change does not alter release profiles or ASLR.

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench daily --only branch/list@many-refs,branch/list-json@many-refs,branch/merged@many-refs,branch/create@many-refs,branch/delete@many-refs,branch/rename@many-refs,branch/list@many-refs-loose,branch/merged-ref@many-refs,tag/list@many-refs,tag/create@many-refs --output /tmp/branch-inventory.json
```

## grit-lib: parallel object hashing

Criterion group `hash/batch_parallel` (`cargo bench -p grit-lib --bench objects hash/batch_parallel`) hashes **20 000** Git blob objects of **16 KiB** each using [`hash_objects_parallel`](https://docs.rs/grit-lib/latest/grit_lib/objects/hash/fn.hash_objects_parallel.html) (SHA-1, release build).

| Threads | Time (median) | Throughput |
| ------- | ------------- | ---------- |
| 1       | 173 ms        | 1.76 GiB/s |
| 8       | 44.5 ms       | 6.86 GiB/s |

**Speedup (8 vs 1 thread): ~3.9×** on the factory VM (2026-10-07).

Parallel hashing falls back to a serial loop when there are fewer than [`PAR_HASH_MIN_ITEMS`](https://docs.rs/grit-lib/latest/grit_lib/objects/hash/constant.PAR_HASH_MIN_ITEMS.html) (32) objects or less than [`PAR_HASH_MIN_TOTAL_BYTES`](https://docs.rs/grit-lib/latest/grit_lib/objects/hash/constant.PAR_HASH_MIN_TOTAL_BYTES.html) (256 KiB) of payload, so small batches avoid thread overhead.

## grit-lib: OID hash maps

Criterion bench **`oid_hashmap`** (`cargo bench -p grit-lib --bench oid_hashmap`) compares `std` `HashMap`/`HashSet` keyed by [`ObjectId`](https://docs.rs/grit-lib/latest/grit_lib/objects/struct.ObjectId.html) (SipHash over the full digest) against [`OidHashMap`](https://docs.rs/grit-lib/latest/grit_lib/objects/type.OidHashMap.html) / [`OidHashSet`](https://docs.rs/grit-lib/latest/grit_lib/objects/type.OidHashSet.html) (hashbrown tables with [`OidBuildHasher`](https://docs.rs/grit-lib/latest/grit_lib/objects/struct.OidBuildHasher.html): first eight digest bytes; [`finish`](https://docs.rs/grit-lib/latest/grit_lib/objects/struct.OidHasher.html#method.finish) XORs that prefix with a process-wide random value). Insert has two symmetric variants at **1 M** keys (SHA-1 OIDs from seeded `ObjectHasher`, distinct first-eight-byte prefixes): **`oid_hashmap_insert_presized`** (both sides call `with_capacity`) and **`oid_hashmap_insert_default_growth`** (both start from empty maps). Factory VM medians from `cargo bench -p grit-lib --bench oid_hashmap -- --sample-size 10 --warm-up-time 1 --measurement-time 3 oid_hashmap_insert` (2026-10-21): presized **~72 ms** SipHash vs **~49 ms** OID (~**1.48×**); default-growth **~115 ms** vs **~72 ms** (~**1.59×**). Lookup and contains groups use the same pre-built tables on both sides. Hot paths (rev-list commit cache, pack delta planning, fsck connectivity, ODB session caches, and related maps) call `OidHashMap::with_capacity_and_hasher` / `OidHashSet::with_capacity_and_hasher` when an object count is known; public APIs still expose standard `HashSet`/`HashMap` at boundaries.

## Pluggable object store (0.6 sign-off, 2026-11-08)

Headline census on `dist` builds in one sitting (12–20 hyperfine samples per row, `--offline`). **Base** grit is `f52021f4cb` (last commit before [`RepositoryOptions::with_object_store`](https://docs.rs/grit-lib/latest/grit_lib/config/environment/struct.RepositoryOptions.html#method.with_object_store), parent of the pluggable-store plan). **After** is `90f552da7e` on the same VM with `--baseline-grit` timing the base binary on every headline row. Reports: `pluggable-store-7249-before.json` → `pluggable-store-7249-after.json`.

**Sign-off (2026-11-08 rerun, after JSON `2026-11-08T17:53:51.439278834Z`):** After census on this VM with `GRIT_BENCH_DROP_CACHES=1`, fail-closed `vm_hygiene`, and **interleaved** `git` → after `grit` → `--baseline-grit` timed samples in one sitting (`dist` grit-bench from this branch; grit **`90f552da7e`** vs **`f52021f4cb`**). Post-hoc baseline-only timing inflated peak-RSS deltas (e.g. **`status/clean@synth-10k`** looked +20%); interleaving fixes that comparison. Peak RSS vs interleaved `baseline_grit` on the ±2% gate rows in this report: **`diff/worktree@synth-10k`** +1.1%, **`stash/push-pop@synth-10k`** +1.9%, **`switch/branch@synth-10k`** +1.9%, **`status/clean@synth-10k`** +4.0%, **`merge/clean@synth-10k`** +3.3% (vs the before-census grit column: **`merge/clean@synth-10k`** +1.4%). Dedicated reruns of the five gate rows on the same VM with the same binaries land within ±2% except merge (typical +2.1–3.0%). Criterion matrix (`odb_read`, `odb_backend`, …) is still outstanding.

| Headline | git ms / MB | base ms / MB | after ms / MB | after vs base time | after vs base RSS |
| --- | ---: | ---: | ---: | ---: | ---: |
| `add/modified@synth-10k` | 16.00 / 11.8 | 9.27 / 17.8 | 9.36 / 18.1 | +0.9% | +1.2% |
| `blame/file@git-git` | 266.32 / 232.8 | 216.75 / 253.6 | 211.62 / 253.7 | -2.4% | +0.0% |
| `branch/list-json@many-refs` | 1.53 / 7.6 | 1.92 / 9.9 | 2.08 / 10.0 | +8.7% | +0.1% |
| `branch/list@many-refs` | 1.28 / 6.4 | 1.86 / 9.5 | 1.88 / 9.5 | +1.1% | +0.5% |
| `branch/merged@many-refs` | 1.42 / 7.3 | 1.96 / 10.2 | 1.99 / 10.2 | +1.8% | +0.5% |
| `clone/local@synth-10k` | 931.70 / 35.4 | 805.36 / 28.2 | 811.17 / 28.3 | +0.7% | +0.4% |
| `commit/small@synth-10k` | 25.95 / 13.3 | 15.38 / 15.8 | 15.28 / 15.8 | -0.7% | +0.5% |
| `diff/worktree@synth-10k` | 6.49 / 9.4 | 5.88 / 12.8 | 5.98 / 13.0 | +1.6% | +1.1% |
| `fetch/incremental@git-git` | 107.56 / 67.6 | 51.77 / 43.8 | 50.92 / 43.4 | -1.6% | -0.9% |
| `log/json-100@git-git` | 2.98 / 14.2 | 3.80 / 17.9 | 3.91 / 17.7 | +2.7% | -1.3% |
| `log/oneline-100@git-git` | 2.87 / 14.2 | 3.56 / 17.3 | 3.60 / 17.6 | +1.0% | +1.8% |
| `merge/clean@synth-10k` | 96.48 / 13.5 | 49.91 / 16.8 | 50.00 / 17.4 | +0.2% | +3.3% |
| `pick/one@synth-10k` | 110.15 / 8.9 | 48.74 / 16.4 | 48.51 / 17.2 | -0.5% | +4.7% |
| `pull/fast-forward@git-git` | 116.96 / 73.0 | 56.74 / 44.0 | 56.64 / 44.1 | -0.2% | +0.2% |
| `pull/rebase@synth-10k` | 69.12 / 39.9 | 17.95 / 22.9 | 17.81 / 22.1 | -0.8% | -3.2% |
| `push/force-with-lease@git-git` | 2083.43 / 318.5 | 933.19 / 68.2 | 931.76 / 68.2 | -0.2% | -0.1% |
| `push/incremental@git-git` | 2115.81 / 318.4 | 1396.43 / 84.6 | 1395.90 / 84.3 | -0.0% | -0.4% |
| `show/head@git-git` | 1.84 / 8.7 | 2.31 / 9.7 | 2.26 / 10.0 | -1.9% | +3.2% |
| `stash/push-pop@synth-10k` | 188.00 / 17.6 | 112.96 / 18.7 | 112.13 / 19.1 | -0.7% | +1.9% |
| `status/clean@synth-10k` | 6.57 / 9.1 | 6.36 / 12.1 | 6.38 / 12.6 | +0.3% | +4.0% |
| `status/dirty@synth-10k` | 9.92 / 9.3 | 6.81 / 12.3 | 6.88 / 13.0 | +1.0% | +5.6% |
| `switch/branch@synth-10k` | 20.83 / 10.2 | 15.05 / 14.0 | 15.22 / 14.3 | +1.1% | +1.9% |

```bash
cargo build --profile dist -p grit-cli -p grit-utils -p grit-http-server
# Base worktree at f52021f4cb03fe824be5bc766baf4f543979d739
./target/dist/grit-bench fixtures prepare --only many-refs,git-git,git-git-wt,synth-10k,synth-10k-smoke
GRIT_BENCH_REPORT_GRIT_COMMIT=f52021f4c ./target/dist/grit-bench daily --offline --min-runs 12 \
  --only '<22 headline ids>' --grit /path/to/base/target/dist/grit \
  --format json --output grit-utils/baselines/pluggable-store-7249-before.json
GRIT_BENCH_DROP_CACHES=1 GRIT_BENCH_REPORT_GRIT_COMMIT=$(git rev-parse HEAD) ./target/dist/grit-bench daily --offline --min-runs 12 \
  --only '<22 headline ids>' --grit ./target/dist/grit \
  --baseline-grit /path/to/base/target/dist/grit \
  --format json --output grit-utils/baselines/pluggable-store-7249-after.json
```

## grit-lib: ODB backend (pluggable object store)

Criterion group **`odb_backend`** compares post-refactor [`Odb`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/struct.Odb.html) against the saved baseline **`odb-before`** (`cargo bench -p grit-lib --bench odb_backend -- --baseline odb-before`). Built-in layouts route hot reads through [`FilesSource`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/store/struct.FilesSource.html) on the primary store and [`CompositeStore`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/store/struct.CompositeStore.html) for alternates so behaviour matches the pre-refactor oracle tests while keeping a pluggable [`ObjectStore`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/trait.ObjectStore.html) boundary.

**Earlier capture (factory VM, 2026-10-09):** Command-level **`grit-bench odb-backend`** baselines are in **`odb-backend-before.json`** and **`odb-backend-after.json`** (pre/post first ODB refactor). Re-run Criterion with `--baseline odb-before` when hot paths change; the 0.6 sign-off headline census is in **`pluggable-store-7249-*.json`** above.

Criterion measures [`Odb::read`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/struct.Odb.html), [`read_info`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/struct.Odb.html#method.read_info), [`exists`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/struct.Odb.html#method.exists) (hit and miss), and [`write`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/struct.Odb.html#method.write) on four deterministic layouts built in `grit-lib/benches/fixture.rs`:

| Fixture | Shape |
| --- | --- |
| `loose_10k` | 10 000 loose blobs, no packs |
| `pack_100k` | Single repacked pack (~100 000 objects) |
| `midx_8` | Eight pack layers with multi-pack-index (`git prune-packed` so hits are pack-only, not loose) |
| `alternate_only` | Primary store empty; objects only in an alternate |

An additional benchmark, **`odb_backend/exists_miss_loop_10k`**, calls `exists` on 10 000 distinct missing OIDs against the 100k pack fixture (models add/status miss scans).

## grit-lib: abbreviating ids on custom object stores

Criterion bench **`abbreviate`** (`cargo bench -p grit-lib --bench abbreviate`) calls [`RevWalk::abbreviate`](https://docs.rs/grit-lib/latest/grit_lib/revwalk/handle/struct.RevWalk.html#method.abbreviate) with a 7-digit minimum for 200 ids out of 200,000 blobs held in three stores: a [`MemoryStore`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/store/struct.MemoryStore.html), a store that implements only the required [`ObjectStore`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/store/trait.ObjectStore.html) methods (so it uses the default full-scan `lookup_prefix`), and one pack in a files ODB. Factory VM medians (2026-10-24, issue #1736):

| Store | Before | After |
| --- | --- | --- |
| `MemoryStore` | 3.89 s | 168 ms |
| Required methods only | 3.90 s | 229 ms |
| Files ODB, one pack | 668 µs | 601 µs |

Prefix matching now compares raw id bytes against a prefix parsed once ([`ObjectIdPrefix`](https://docs.rs/grit-lib/latest/grit_lib/objects/struct.ObjectIdPrefix.html)) instead of formatting every candidate as hex, and abbreviation makes one lookup per id instead of one per candidate length. A store without an index still scans every id once per lookup; override `lookup_prefix` with a range scan to avoid that (see [Object stores](https://grit-scm.com/docs/library/object-stores/index.md#short-ids-and-prefix-lookup)).

## Incremental push over SSH

The send-pack walk stops when no queued commit remains reachable only from the wanted tips. Advertised commits still propagate exclusion flags through shared ancestry; unrelated old refs no longer force a walk to the roots. Regression tests bound actual commit reads with a direct-parent have plus a near-root have, compare merge-history packs with system Git, and push through Git's receive-pack with annotated blob/tree tags advertised.

Factory VM, 2026-10-26, 4 vCPU, Rust 1.99, `release` profile, system Git 2.55.0: one new commit over a local fake-SSH wrapper on upstream `git/git` at `7e34d8174dcda3435d1404c9a30bef46afe6c13e` (82,563 commits on master). The source and bare receiver retain upstream refs and tags and use shared object storage without bitmaps. Before each run, reset the remote branch to the new commit's parent; push `HEAD:refs/heads/master` with each client. One warmup followed by six interleaved runs uses Python `perf_counter` wall time and `/usr/bin/time -f %M` peak RSS, with compilation paused. The committed samples are in `grit-utils/baselines/push-one-commit-ssh.json`.

| Client | Median wall time | Peak RSS |
| --- | ---: | ---: |
| System Git | 71.3 ms | 47.3 MiB |
| Grit before the cutoff fix (`c190a4fff`) | 2,629.2 ms | 74.6 MiB |
| Grit after | 144.7 ms | 51.9 MiB |

The cutoff reduces median push time by **18.2×**, with about **30% lower peak RSS**; Grit remains about **2.0×** slower than Git on this fixture. Every push updated the expected ref. `git fsck --strict --no-dangling` succeeded on the receiver, with warnings for historical upstream tag/tree objects. The small SSH regression fixture passes strict fsck without those historical objects.

### One new commit over ssh, git:// and smart HTTP (2026-10-27, #1975)

After the walk cutoff, the cost of a one-commit push still grew with the number of branches the remote advertises. Three changes remove that cost:

- **Thin-pack bases.** As with Git's `--thin`, these now come only from the edge commits where the walk met the remote's history. Before, they came from every advertised branch, and delta planning walked each of those trees once for on-disk delta reuse and again for every thin target.
- **Advertised tips.** Each tip is read once while its tags are followed.
- **Bitmaps.** Opening a pack bitmap no longer reads and checksums every `.idx` when the pack has no `.bitmap`.

Two smaller fixes follow. Pack placement and delta reuse look up a few objects instead of scanning the whole pack index. The git:// and ssh readers are buffered, so the advertisement takes a handful of reads instead of two per ref.

Factory VM, 4 vCPU Xeon, Linux 6.12, Rust 1.99, `release` profile, system Git 2.43.0. The fixture is a `git clone --bare` of upstream `git/git` (82k commits, about 1,000 refs), reached through a fake-ssh wrapper, `git daemon --enable=receive-pack` and `grit-http-server`. The worktree is a `grit clone` over git://. Each push sends a new branch holding one commit that appends a line to `README.md`. One warmup is followed by 9 runs per client, interleaved across clients and transports. Wall time is Python `perf_counter`; RSS is the client's `wait4` peak. Every push adds a branch to the remote, which grew from 137 to 257 branches during the run. Main's time grows with the branch count (0.17 s on a fresh clone); this branch's does not. Samples are in `grit-utils/baselines/push-one-commit-1975.json`.

| Transport | System Git | Grit 2a0f003c1 | Grit main (b3a18044a) | Grit after |
| --- | ---: | ---: | ---: | ---: |
| fake SSH | 64.6 ms / 43.7 MiB | 1,833.6 ms / 93.6 MiB | 437.9 ms / 58.2 MiB | 68.3 ms / 43.7 MiB |
| git daemon (git://) | 86.4 ms / 17.7 MiB | 1,854.4 ms / 93.3 MiB | 485.9 ms / 57.9 MiB | 132.5 ms / 33.8 MiB |
| grit-http-server smart HTTP | 49.8 ms / 17.6 MiB | 1,785.2 ms / 94.9 MiB | 467.4 ms / 59.2 MiB | 45.3 ms / 35.0 MiB |

The branch matches Git over ssh and smart HTTP. Over ssh, peak RSS includes the server's `git receive-pack`, since the fake-ssh wrapper runs it as a child of the client.

The dashboard's `push/incremental`, `push/incremental-http` and `push/force-with-lease` scenarios on `git-git` (`daily-push-1975-before.json` / `-after.json`) improve by 6–9% in median wall time and 1–6% in peak RSS. Git's own time dominates those scenarios.

One known gap remains: `grit push` lists the remote's refs over a separate upload-pack connection before it connects to receive-pack. Git uses one connection. On git:// that second connection accounts for most of the remaining difference.

## Whole-repository push over smart HTTP

Pushing all of a repository to an empty remote is where a client holds the most pack data. `grit push` used to build the whole pack in memory, copy it into the HTTP request body, and map and index every source pack twice (once for the object walk and again for the delta reuse scan). Now the pack streams into the request while it is written: a receive-pack body larger than 1 MiB goes out with chunked transfer encoding after a small probe request, as Git does. The whole build shares one mapping of each source pack, and reused objects and deltas are written straight from that mapping.

Factory VM, 4 vCPU, Linux 6.12, Rust 1.99, `release` profile, system Git 2.43.0. `grit push` and `git push` of `master` from a clone of rust-lang/cargo (23.6k commits, 195k objects, 80 MB pack) to fresh bare repositories served by one `grit-http-server` on localhost. Five interleaved runs per client, `/usr/bin/time` wall time and peak RSS:

| Client | Wall time | Peak RSS |
| --- | --- | --- |
| `git push` | 2.38–2.61 s | 163 MB |
| `grit push`, before | 3.58–3.92 s | 427 MB |
| `grit push`, after | 3.34–3.53 s | 268 MB |

A pack of four incompressible 16 MiB blobs shows what still grows with the pack: git peaks at 71 MB, grit at 213 MB before and 84 MB after, the rest being the mapped pack both tools read. Every pushed repository passes `git fsck --strict`. Most of the remaining time difference is the object walk (about 1.0 s against 0.41 s for `git rev-list --objects`).

The gate keeps this from regressing: `grit-utils/tests/push_http_cargo_shape_budget.rs` pushes a `CargoShape::SMOKE` history and the four-blob repository with both clients to one stock `git http-backend` server, and requires grit's peak memory above its `log -n 1` baseline to stay within 2× of git's (1.5× for the blobs), and within 2× of git's time in optimized builds.

## grit-lib: pack write (100k reuse vs fresh)

Criterion group **`pack`** (`cargo bench -p grit-lib --bench pack_write`) rebuilds the full closure of the repacked **`pack_100k`** fixture (~100 000 objects) via the library's parallel pack builder:

| Benchmark | What it measures |
| --- | --- |
| **`write_100k_reuse`** | Delta planning with on-disk delta and whole-object reuse (`PackBuildOptions::for_local_copy`, bitmaps off) |
| **`write_100k_fresh`** | Same window/depth but **`reuse_objects`** and **`reuse_deltas`** disabled (recompress and re-delta everything) |

Throughput is reported in objects per second using the **emitted** pack header count from a probe rebuild of the full closure from the fixture **history tip** (cached repo `HEAD` or the generated repo tip), not the midpoint `hit` OID used for ODB read micro-benchmarks. Each iteration asserts that count matches the probe. Populate **`GRIT_BENCH_ODB_CACHE`** (or run **`grit-bench odb-backend`** once) so the fixture reuses the cached repacked repo instead of rebuilding 100k commits inline.

```bash
cargo bench -p grit-lib --bench pack_write -- write_100k
```

Command-level ODB scenarios on the repacked **100k-file / 1000-commit** synthetic repo live in **`grit-utils/baselines/odb-backend-before.json`** and **`grit-utils/baselines/odb-backend-after.json`** (`grit-bench odb-backend`): library drivers (`grit-bench drive …`) vs **`git cat-file --batch`**, **`git cat-file --batch-check`**, and **`git rev-list --objects --all`** on the same sorted OID stdin list where applicable.

```bash
cargo build --release -p grit-utils
./target/release/grit-bench odb-backend --format json --output grit-utils/baselines/odb-backend-after.json
cargo bench -p grit-lib --bench odb_backend -- --baseline odb-before
make docs
```

Run `grit-bench odb-backend` once before the Criterion command (or leave a populated `GRIT_BENCH_ODB_CACHE`) so the `pack_100k` fixture reuses the repacked 100k repo instead of rebuilding 100k commits inline.

**Recorded on the factory VM (2026-10-09, repacked 100k synthetic repo).** Pre-refactor numbers are in **`odb-backend-before.json`**; post-refactor in **`odb-backend-after.json`** (also drives the generated ODB rows below).

| Scenario | Git median | Grit median (before → after) | Grit / Git (before → after) |
| --- | ---: | ---: | ---: |
| `cat-file-batch-hot-path-100k` | ~144 ms | 1 117 ms → ~356 ms | **7.4× → ~2.5×** |
| `cat-file-batch-check-hot-path-100k` | ~52 ms | 1 261 ms → ~239 ms | **24.2× → ~5.2×** |
| `rev-list-objects-odb-backend-hot-path-100k` | ~65 ms | 6 916 ms → ~6 189 ms | **99.1× → ~96×** |

Rev-list **ratios** can rise when system **`git`** speeds up between captures even if Grit wall time improves; compare absolute Grit medians when judging regressions.

The rev-list gap is dominated by history/object enumeration in `grit-lib`, not bulk pack I/O; cat-file scenarios exercise ODB read and `read_info` paths directly.

## grit-lib: parallel packed ODB reads

Criterion group **`odb_read_parallel/packed_blobs`** (`cargo bench -p grit-lib --bench odb_read_parallel`) measures throughput when worker threads share one [`Odb`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/struct.Odb.html) and hold [`Odb::with_pack_read_context`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/struct.Odb.html#method.with_pack_read_context) for the worker lifetime. Each iteration reads a partition of **50 000** whole-object pack blobs (**3072-byte** payload; loose copies are removed after packing so hits are pack-only). Baseline **`odb-read-parallel.json`** records Criterion central estimates on the factory VM (**4 logical CPUs**, Rust **1.99** stable).

| Threads | Median time (50k reads) | Throughput |
| ------- | ----------------------: | ---------- |
| 1 | ~149 ms | ~336 Kelem/s |
| 4 | ~51 ms | ~983 Kelem/s |

**Speedup (4 vs 1 thread): ~2.9×** on the factory VM (2026-10-11). This exercises the explicit batch read-context API and shared [`PackStore`](https://docs.rs/grit-lib/latest/grit_lib/objects/pack_store/struct.PackStore.html) catalog, not ad-hoc per-read context installation.

```bash
cargo bench -p grit-lib --bench odb_read_parallel -- packed_blobs
```

## grit-bench: reachability bitmaps (`git.git`)

Scenario group **`bitmaps`** measures three acceptance metrics on upstream **`git/git`**: commit counting, object counting with bitmap index, and stateless **`upload-pack`** clone serving. A fourth scenario, **`serve-clone-no-bitmaps-git.git`**, serves the same clone from a copy of the fixture with the `.bitmap` removed (one Git-written pack with its `.rev` reverse index, `ensure_git_git_no_bitmaps`), which is how most servers that never wrote bitmaps look: objects are enumerated by walking history and the pack reuses the on-disk deltas. **`serve-clone-blobless-git.git`** and **`serve-clone-blobless-no-bitmaps-git.git`** send the same request with `filter blob:none` (the server side of a blobless partial clone, `drive serve-blobless`), with `uploadpack.allowFilter` enabled on both fixtures. The fixture is a bare **single-branch tag clone** at **`v2.47.0`** (override with **`GRIT_BENCH_GIT_GIT_TAG`**) under **`GRIT_BENCH_ODB_CACHE`**, prepared once with **Git-written** pack bitmaps (`git repack -adb`, `git commit-graph write --reachable`) and separately with **grit-written** pack bitmaps (`git repack -ad`, commit-graph, then `PackBitmapWriter` — see `ensure_bitmaps_grit_repo` in `grit-utils/src/bitmap_fixture.rs`). Library workloads use hidden **`grit-bench drive`** commands (`rev-list-count`, `rev-list-count-objects`, `serve-clone`); Git runs the equivalent **`git rev-list`** / **`git upload-pack`** invocations with the same isolated config as other grit-bench suites.

**Methodology.** Hyperfine warmups and minimum runs match other factory captures (see each JSON `machine` block). Ratios are **Grit median ÷ Git median**. **Peak RSS** is recorded per side via `grit-bench measure-rss` / hyperfine wrappers. **`serve-clone`** runs under **`GRIT_BENCH_SERVE_TIMEOUT_SECS`** (default **120**) and **`GRIT_BENCH_SERVE_MAX_RSS_GIB`** (default **12**). Reproducibility: run the suite twice back-to-back on the same fixture; medians should agree within **±10% relative** (`grit-bench compare … --tolerance 0.10` uses an absolute ratio tolerance — interpret large ratios accordingly).

```bash
cargo build --release -p grit-utils -p grit-cli
./target/release/grit-bench bitmaps --format json --output grit-utils/baselines/bitmaps-after.json
# Optional grit-written pack bitmap sidecar (separate JSON; default fixture stays Git-prepared):
GRIT_BENCH_BITMAP_FIXTURE=grit ./target/release/grit-bench bitmaps --format json --output grit-utils/baselines/bitmaps-grit-written.json
```

**Before (`bitmaps-before.json`, 2026-10-10, pre-reader/writer stack):** bitmap reader missing — object count ~**23 000×** vs Git; serve-clone hit **memory_cap**.

**First acceptance run (factory VM 2026-10-16, release `grit` on the pack-reuse + bitmap stack):** `rev-list-count` **18.0×**, `rev-list-count-objects` **63.2×**, and `serve-clone-git.git` hit the **120 s** cap (`exit 124`). Two back-to-back runs agreed within **~6–8% relative** on Grit medians for the rev-list scenarios.

**Current (`bitmaps-after.json`, 4-vCPU Xeon VM 2026-10-25, Git 2.43.0, after the filtered-serve fix for [#1819](https://github.com/gitbutlerapp/grit/issues/1819)):**

| Scenario | Git median | Grit median | Ratio | Git peak RSS | Grit peak RSS |
| --- | ---: | ---: | ---: | ---: | ---: |
| `rev-list-count-git.git` | 56.5 ms | 686.0 ms | **12.1×** | ~29 MiB | ~79 MiB |
| `rev-list-count-objects-git.git` | 16.9 ms | 719.7 ms | **42.7×** | ~23 MiB | ~79 MiB |
| `serve-clone-git.git` | 654.4 ms | 4235.8 ms | **6.5×** | ~278 MiB | ~765 MiB |
| `serve-clone-blobless-git.git` | 296.1 ms | 2146.1 ms | **7.2×** | ~156 MiB | ~364 MiB |
| `serve-clone-no-bitmaps-git.git` | 3920.3 ms | 12079.7 ms | **3.1×** | ~462 MiB | ~811 MiB |
| `serve-clone-blobless-no-bitmaps-git.git` | 2810.2 ms | 8817.5 ms | **3.1×** | ~325 MiB | ~446 MiB |

Serving used to load every object of the clone into memory (blobs included) and decode the pack's `.rev` file again for every `OFS_DELTA` base, so both serve scenarios ran past the cap. Upload-pack now copies packed objects and deltas as they are, delta-searches only blobs that are not already deltas, and streams the pack. On `rust-lang/cargo` (195k objects, 80 MB pack, no bitmap) the server side of a full clone went from **91.3 s / 3.6 GB** peak RSS to **3.0 s / 420 MB** (Git **2.5 s / 161 MB**).

**Filtered serving.** Without a usable bitmap, a `filter` request used to go through a separate `rev-list` walk that read every reachable tree once per commit. The pack enumeration walk now applies the filter itself, so a filtered pack costs what an unfiltered one does minus the objects it leaves out. Server side of `blob:none` requests on the same VM (single runs; the bitmap row is a hyperfine mean of 10):

| Request | Git | Grit before | Grit after |
| --- | ---: | ---: | ---: |
| `git.git` every ref, no bitmap | 2.9 s / 316 MB | 277.9 s / 11.7 GB | 8.6 s / 451 MB |
| `git.git` every ref, bitmap | 0.33 s / 143 MB | 2.18 s / 367 MB | 2.19 s / 367 MB |
| `rust-lang/cargo` clone, no bitmap | 1.4–1.7 s / 148 MB | 55.7 s / 1.8 GB, then failed | 1.4–1.5 s / 260 MB |
| `rust-lang/cargo` fetch of tag `0.58.0` | 0.53 s / 69 MB | 5.8 s, then failed | 0.65 s / 132 MB |
| `rust-lang/cargo` fetch of 400 new commits | 0.16 s / 29 MB | failed | 1.65 s / 161 MB |

The cargo clone and tag fetch used to fail because two annotated tags on one commit (`0.58`, `0.58.0`) kept only one tag object in the pack, and every filtered fetch with haves failed. With a bitmap, the bitmap already selected the objects; the remaining gap there is the same planning and writing cost as the unfiltered `serve-clone-git.git`.

**Grit-written bitmap fixture:** set **`GRIT_BENCH_BITMAP_FIXTURE=grit`** when running `grit-bench bitmaps` to prepare the same pinned **`git.git`** clone with **`git repack -ad`**, commit-graph, and a grit-lib **`PackBitmapWriter`** sidecar (`ensure_bitmaps_grit_repo`). The committed **`bitmaps-after.json`** uses the **Git-prepared** fixture (default); capture a separate JSON for grit-written bitmaps when comparing interoperability (see factory notes for the grit-fixture run ratios).

**Acceptance bar (roadmap item 14):** each ratio **≤ 1.2×** (and serve-clone within the same bar). The after run **passes compatibility** (bitmaps read/write, serve packs validate with `git index-pack --strict` after the thin-pack reuse fix) but **misses the 1.2× performance bar** on all three command-level metrics on this tag-heavy **`git.git`** fixture (~954 tag refs). Follow-ups (not retargeting the bar):

- **Multi-tip bitmap counting:** optimize `rev-list --count --all` / `--objects --use-bitmap-index` when hundreds of tag tips are present (Git ~60 ms vs grit ~1 s on the same fixture).
- **Serve-clone wall time and memory:** both serve scenarios now finish well under the cap. What remains is the history walk that enumerates objects without a bitmap (about 3× Git's on `git.git`, dominated by visited-set probes and tree reads), bitmap-path planning, and memory held by per-object maps during delta-reuse planning.
- **Re-run `bitmaps-after.json`** after those land; keep **`bitmaps-before.json`** for regression context.

Intermediate capture on the same VM before the final writer stack (**2026-10-10**) remains in **`bitmaps-before.json`** for the generated tables (`rev-list-count` **~3.3×**, object count unusable, serve **memory_cap**).

## grit-lib: writing pack bitmaps

`grit gc repack --bitmaps` on an already-packed `rust-lang/cargo` clone (195,008 objects in one 81 MB pack, 24,341 commits, 110,673 trees, 109 release branches) used to spend about 56 s writing the `.bitmap` ([#1877](https://github.com/gitbutlerapp/grit/issues/1877)). The writer read every blob while walking trees, re-read commits inside sorts, walked full reachability a second time to check closure, and read the whole pack to get its 20-byte trailer. It now reads only commits and trees. Types come from pack entry headers, and an unwritten bitmap at each first-parent fork keeps release branches from re-walking the main line below the fork (tree reads went from 388k to about the number of distinct trees). Commit and type bitmaps are byte-identical to before; 66 of 195,008 name-hash entries changed.

4-vCPU VM, Git 2.43.0, wall time and peak RSS (`/usr/bin/time`), each on a fresh copy of the clone:

| Operation | Git | Grit before | Grit after |
| --- | ---: | ---: | ---: |
| `grit gc repack --bitmaps` vs `git repack -a -d -b` | 1.6 s / 174 MB | 58.3 s / 414 MB | 2.9 s / 414 MB |
| `grit repack` vs `git repack -a -d` (no bitmap, for reference) | 1.15 s / 169 MB | 2.0 s / 414 MB | 2.0–2.25 s / 414 MB |
| Criterion `pack_bitmap/write_from_git_bitmap` (bare cargo clone) | — | 54.87 s | 0.967 s |

The peak RSS comes from the repack itself, not from the bitmap write. Reproduce the library number with:

```bash
GRIT_GIT_GIT_BARE=/tmp/cargo.git cargo bench -p grit-lib --bench pack_bitmap -- write_from_git_bitmap
```

## grit-lib: parallel index-pack (in-memory)

Release [`index_pack_bench`](https://github.com/gitbutlerapp/grit/tree/main/grit-lib/examples/index_pack_bench.rs) on a **~112 500-object** depth-50 pack (`git fast-import` + `git repack -adf --depth=50`, factory VM 2026-10-12). Grit uses [`pack_index_records`](https://docs.rs/grit-lib/latest/grit_lib/objects/unpack_objects/fn.pack_index_records.html) (`GRIT_INDEX_PACK_BENCH_PACK` / `GRIT_INDEX_PACK_THREADS`); timings cover in-memory record building only (no `.idx` write). System **`git index-pack`** rows are end-to-end on a fresh temp directory each run (writes `.idx`).

| Command | Median (5 runs) |
| ------- | ---------------- |
| grit index-pack (records only), 1 thread | 1.21 s |
| grit index-pack (records only), 4 threads | 0.49 s |
| `git index-pack --threads=1` (end-to-end) | 1.25 s |
| `git index-pack --threads=4` (end-to-end) | 0.43 s |

**Speedup (4 vs 1 thread, grit records): ~2.47×.** Multi-threaded builds overlap pack scanning with batched parallel inflate+hash (ordered work pool; pack order preserved in the public record list). The single-thread path scans once and inflates+hashes serially—no redundant second skip pass. Delta chains still inflate, apply, and hash in parallel rounds.

Integration coverage uses a **≥ 50 000** object fixture built with `git fast-import` and shared across tests via `OnceLock` (`grit-lib/tests/index_pack_parallel.rs`); the default release suite completes in about **41 s** on the factory VM.

## Hashing

Checksum throughput uses plain accelerated SHA-1/256 from the `sha1` / `sha2` crates ([`HashingWriter`](https://docs.rs/grit-lib/latest/grit_lib/objects/hash/struct.HashingWriter.html), Criterion `hash_sha1` / `hashing_writer_sha1`). Object-id hashing on ingestion uses [`ObjectHasher::for_object`](https://docs.rs/grit-lib/latest/grit_lib/objects/hash/struct.ObjectHasher.html) with the **sha1dc** crate (collision detection on the canonical object header + payload). OpenSSL numbers are from `openssl speed -evp sha1 -bytes 16384` and `openssl speed -evp sha256` (16 KiB blocks where shown). Git blob hashing uses `git hash-object` on a 256 MiB file vs the `gritx-hash-file` example (same canonical blob object id, no object-database write).

**Environment (2026-10-21):** Intel Xeon (SHA-NI present), Linux, `git` 2.43.0, OpenSSL 3.0.13, Rust stable 1.99. `hash-info --json` reports `{"sha1":"x86_sha_ni","sha256":"x86_sha_ni"}`.

### Micro (Criterion + hyperfine)

| Workload | grit-lib | Reference | grit / reference |
| -------- | -------- | --------- | ---------------- |
| SHA-1 checksum, 1 MiB buffer | 1.59 GiB/s | OpenSSL SHA-1 @ 16 KiB: 1.88 GiB/s | **85%** |
| SHA-1 checksum via `HashingWriter`, 1 MiB | 1.93 GiB/s | OpenSSL SHA-1 @ 16 KiB: 1.88 GiB/s | **103%** |
| SHA-256, 1 MiB buffer | 1.76 GiB/s | OpenSSL SHA-256 @ 16 KiB: 1.73 GiB/s | **101%** |
| SHA-1 object id (`hash_object_sha1`), 1 MiB | 1.57 GiB/s | effective `git hash-object` throughput on 256 MiB: ~0.40 GiB/s | **~3.9× throughput** |
| SHA-1 blob, 256 MiB file (hyperfine) | 360 ms | `git hash-object`: 612 ms | **1.70× faster** |

Plain SHA-1 checksums at 1 MiB meet the project bar (≥ 95% of OpenSSL on the same CPU) on the `HashingWriter` path used for pack checksums. SHA-256 meets ≥ 90% of OpenSSL. Object-id hashing keeps collision detection (sha1dc) while staying well above stock Git’s sha1dc-backed `hash-object` throughput on large blobs (see **`hash-after.json`** `hash-object-256mib`: grit/git median ratio **0.61×**, **1.64×** faster than system git on the same 256 MiB `hash-object -w` run).

### End-to-end (`hash-after.json`, factory VM 2026-10-21)

Committed baseline **`grit-utils/baselines/hash-after.json`** (paired with **`hash-before.json`**) from `grit-bench hash`: 256 MiB `hash-object -w`, repacked 100k-object **`index-pack`** (SHAttered blobs excluded so grit and git compare fairly), and medium **`fsck`**.

| Scenario | Grit median | Git median | Grit / Git |
| -------- | ----------- | ---------- | ---------- |
| `hash-object-256mib` | 3.819 s | 6.263 s | **0.610×** (1.64× faster) |
| `index-pack-git-git` (100k repacked) | 123.1 ms | 197.7 ms | **0.623×** (1.61× faster) |
| `fsck-medium` | 416.7 ms | 239.4 ms | **1.741×** (grit slower; follow-up [#1475](https://maint.grit-scm.com/factory/1475)) |

Versus **`hash-before.json`**, grit median on `hash-object-256mib` is **−3.9%** (3971.9 ms → 3818.5 ms), within the **5%** acceptance bar. Profiling the `hash-object -w` path showed a redundant **payload-only sha1dc pass** on every large blob after canonical object-id hashing; **`grit-lib`** now skips that pass when the payload is larger than 64 KiB and does not start with a SHAttered block prefix. Loose writes use [`hash_object`](https://docs.rs/grit-lib/latest/grit_lib/objects/hash/fn.hash_object.html) for the object id (no duplicate `ObjectHasher` in the store layer). `index-pack` uses the repacked 100k fixture (see **`factory/checked-object-hash-2718`**) so before/after medians for that scenario are not directly comparable.

### Commands

```bash
# Backends selected on this CPU
cargo run -p grit-examples --bin hash-info -- --json

# Criterion throughput (64 B … 64 MiB)
cargo bench -p grit-lib --bench hash

# OpenSSL + hyperfine vs git (creates a 256 MiB temp file)
./scripts/bench-hash.sh
```

## Merge, pick, revert conflict workflows (2026-10-23)

The **`conflicts`** grit-bench group times full porcelain workflows on synthetic **1k** and **10k** file fixtures (reset to a fixed state each hyperfine iteration). Scenarios: clean **`merge --no-ff`**, **`merge` → `restore --theirs` → `add` → `--continue`** on overlapping edits, **`pick`** of twenty commits in one invocation vs `git cherry-pick base..topic`, and **`revert HEAD`** / **`revert --mainline 1`** on a recorded merge commit. In-memory **`merge --check`** stays in **`merge-check.json`** (library driver vs `git merge-tree --write-tree`).

Committed baseline: **`grit-utils/baselines/conflicts.json`**. Factory VM medians vs **`git 2.43.0`** (2026-10-23): **`merge-no-ff`**, **`revert-single`**, and **`revert-mainline`** are at or below Git time at both tiers; **`pick-range-20`** is **~1.15×** at 10k files; **`merge-conflict-continue`** and **`pick-range-20`** at 1k remain above **1.0×** (tracked as speed follow-ups). At **10k** files, **`merge-conflict-continue`** is still **~5.5×** Git — dominated by grit-side restore/add on the wide conflict set, not by merge prediction.

```bash
./target/release/grit-bench conflicts --sizes 1000,10000 --format json --output grit-utils/baselines/conflicts.json
```

## Full-history walks use the commit-graph (2026-10-24)

Listing every commit with a valid commit-graph took as long as walking without one ([#1745](https://github.com/gitbutlerapp/grit/issues/1745)). The walk took parents and dates from the graph, but still inflated every commit object to cache its subject line, and looked each commit up in the graph three times. Commits found in the graph are no longer read during the walk. Callers that print subjects read them for the listed commits only. Each commit now costs one graph lookup.

[`RevWalk::rev_list`](https://docs.rs/grit-lib/latest/grit_lib/revwalk/handle/struct.RevWalk.html#method.rev_list) on upstream `git/git` (82 563 commits, commit-graph from `git gc`), in process, release build, `origin/main` → this change, against system `git 2.43.0`:

| Walk | Git ms | Grit ms with graph, before → after | Grit ms without graph |
| --- | ---: | ---: | ---: |
| default order | 71–106 | 760–860 → **105** | 740 |
| `--topo-order` | 90–126 | 750–760 → **102** | 705 |
| `--first-parent` | 20–30 | 176 → **23** | 167 |

Git's times include process start-up. Criterion `history` groups on the 10k-commit synthetic fixture, `--baseline main`: `revwalk_default`, `revwalk_topo` and `revwalk_date_order` with a graph are **−80 to −82%**, `revwalk/log_limit` and `revwalk/limit_near_full_history` are **−81 to −89%**, and the no-graph rows are within the run-to-run drift of an A/A run on main (±4%). `grit revs HEAD` on `git/git` drops from **1.73 to 1.00 s**. Its peak RSS stays at **283 MB**, because it builds every row (full message included) before rendering. `git log` with subjects takes 0.70 s and 66 MB.

The `revwalk` suite gained `revs-list-deep-no-graph` and `revs-list-deep-commit-graph` (every commit id from `HEAD`, library walk, vs `git rev-list HEAD`) in **`revwalk-queries.json`**:

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench revwalk --format json --output grit-utils/baselines/revwalk-queries.json
cargo bench -p grit-lib --bench history -- 'revwalk_(default|topo|date_order)/'
```

## Big files are streamed (2026-10-25)

`grit add` of a 1 GB file used to read it into memory and copy it twice more on the way to the object store, peaking at 2.4–3× the file size (issue #1850). Files of at least `core.bigFileThreshold` (512 MiB by default) that need no clean conversion are now hashed and deflated in 128 KiB chunks straight into their own pack, as Git does, at the pack compression level. Smaller files lost two copies (the clean conversion takes ownership of the bytes, and `add` deflates straight into the loose object file instead of a buffer kept until the batch finished), so they now peak at about 1× the file size, like Git.

Same Linux VM sitting (4 vCPU), release builds, system `git 2.43.0`, `/usr/bin/time -v`, one run each; grit is `origin/main` before → after:

| Command | Git | Grit before → after (wall) | Peak RSS Git, Grit before → after | `.git` size Git, Grit before → after |
| --- | ---: | ---: | ---: | ---: |
| `add` 1.09 GB text (`seq 1 120000000`) | 19.1 s | 5.3 → **7.4** s | 5 MB, 2,596 → **14** MB | 265, 466 → **271** MB |
| `commit` (auto-stage) 1 GB random | 22.4 s | 12.6 → **18.7** s | 5 MB, 2,997 → **16** MB | 1,000, 1,055 → **1,000** MB |
| `status` after touching the 1 GB file | 19.3 s | 1.9 → **0.8** s | 5 MB, 1,966 → **15** MB | |
| `add` 349 MB text | 3.0 s | 1.7 → **1.3** s | 346, 846 → **355** MB | 90, 155 → 155 MB |
| `add` 300 MB random | 6.3 s | 3.7 → **3.3** s | 298, 909 → **307** MB | 300, 316 → 316 MB |

Big files are slower than before because the pack level (`pack.compression`, default 6) replaces the loose level 1, whose zlib-rs fast path compresses text far worse. Both are still faster than Git, and the repository is now about Git's size. Set `pack.compression` (or `core.compression`) lower to trade size for speed. Files below the threshold keep the loose level, which is why their repository size is unchanged. `grit-bench add-initial` (20,000 small files) and `add` / `status` at 10,000 files are unchanged within noise.

`grit-bench add-big-file` (committed as `add-big-file.json`) times `add` of one untracked text file per size with the index and objects reset between runs, and records peak RSS:

```bash
./target/release/grit-bench --warmup 1 --min-runs 3 --format json --output grit-utils/baselines/add-big-file.json add-big-file --megabytes 256,1024
```

## Large-blob reads and restaging (2026-10-26)

Issue #2010 exposed copies of blob payloads in `cat`, unnecessary writes when
restaging unchanged content, and a local-clone statistics walk that inflated
blobs just to count them. Cat now uses the library's replacement-aware payload
stream. Restaging uses the same streaming hash threshold as status and stores
content only when its object id changes. Local transfers link loose objects,
and the statistics walk reads blob headers without loading their payloads.

Linux, 4 vCPUs, release builds, system Git 2.55.0, baseline `origin/main`
`c190a4fff`. The committed `large-blob-memory-before.json` and
`large-blob-memory-after.json` record `/usr/bin/time -v` wall time and peak RSS
for incompressible binary files. Cat, restore and clone use three runs with
rotating tool order on identical sources; the table reports medians. Status and
add are single runs after moving the file's mtime into the future without
changing its bytes. Every clone passes system `git fsck --strict` and a blob
hash comparison.

This capture predates the related #1849 integration. Current main also contains
the matching human-mode cat stream and loose-clone optimizations; this change
reuses that display API and keeps the restaging, structured-output bounds,
pack serialization and regression coverage described here.

After rebasing onto that integration, `grit cat HEAD:big.bin --json` on the
1 GiB packed fixture returned 173 bytes of metadata at 13.4 MiB peak RSS.
The RSS guard also covers JSON and Markdown output.

| Operation | Git seconds | Main seconds | After seconds | Git RSS MiB | Main RSS MiB | After RSS MiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| Status, 128 MiB | 0.36 | 0.13 | 0.09 | 133.1 | 13.7 | 14.8 |
| Add, 128 MiB | 0.39 | 0.35 | 0.09 | 132.9 | 140.4 | 13.6 |
| Status, 1 GiB | 2.56 | 1.61 | 1.03 | 5.3 | 13.4 | 15.0 |
| Add, 1 GiB | 44.90 | 31.29 | 0.84 | 5.0 | 12.9 | 13.3 |
| Cat, loose 1 GiB | 0.41 | 1.12 | 0.21 | 1028.8 | 4449.7 | 12.7 |
| Restore, loose 1 GiB | 0.73 | 1.02 | 0.94 | 1029.1 | 14.3 | 14.4 |
| Clone, loose 1 GiB | 0.74 | 0.72 | 0.32 | 1029.5 | 2061.0 | 15.4 |
| Cat, packed 1 GiB | 0.37 | 1.33 | 0.23 | 1028.9 | 5474.2 | 1036.8 |
| Restore, packed 1 GiB | 0.69 | 0.99 | 0.94 | 1029.6 | 1038.4 | 1038.4 |
| Clone, packed 1 GiB | 0.71 | 0.38 | 0.31 | 1029.7 | 1039.7 | 1039.8 |

Linux RSS includes resident pages of memory-mapped object files, which explains
the roughly 1 GiB floor for packed reads here. Loose-object cat and clone now
stay near their fixed process overhead. Pack deltas still require reconstruction
in memory. Restore was already streaming on this Linux baseline; these results
do not claim validation of the original Windows-specific restore spike.

Reproduce the memory guard after building both helper binaries:

```bash
cargo build --release -p grit-cli -p grit-utils
GRIT_LARGE_BLOB_MEMORY=1 cargo test -p grit-lib --test integration large_blob_memory::
```

To create the loose 1 GiB fixture, set `core.bigFileThreshold=2g` for the
system `git hash-object -w` command before building its tree and commit. Without
that override, Git stores the large file in a pack. Compare `git cat-file blob
<oid>` with `grit cat HEAD:big.bin`, and local `git clone` with `grit clone`;
redirect cat output to `/dev/null` and use a fresh clone destination each time.

## Clean-tree status (2026-10-24)

Dogfooding on Windows (issue #1702) found `grit status` re-reading tracked files on clean trees and running 3.5–5× slower than Git on rust-lang/cargo and git.git. Four causes, all in `grit-lib`:

- After a stat match, the content-hash skip refused to trust the index when `core.filemode=false` (the Windows default) or when the cached size was zero. Serial scans (fewer than 1000 tracked files, `--threads 1`, `core.preloadIndex=false`) therefore re-hashed every file. Zero size is now trusted only for the empty blob, as in Git.
- `status` never rewrote an index whose entries were racily clean (index written in the same second as the files, as after `git clone` or `git commit`), so every run re-hashed them. It now persists the index once, as Git does, and hashes each racy file only once.
- The untracked walk paid three path probes per directory (`is_dir`, `.git`, `.gitignore`), and the symlinked-ancestor check ran as a serial pre-pass before the parallel stat workers. On Windows each path stat opens a handle. Probes now come from the directory listing or run inside the workers, and Windows uses a single attribute query for them.
- Outside Windows, tracked files are stat-ed by path like Git's preload instead of listing their directories: ext4 hashed `readdir` dominated large directories.

Same Linux VM sitting, release builds, system `git 2.43.0`, clean trees with `GIT_CONFIG_GLOBAL` empty. `git status --porcelain` → `grit status --json`, median ms over 25 runs; grit is `origin/main` before → after:

| Repository | Tracked files | Git | Grit before → after | Tracked part (`-uno` / `--untracked no`): Git, Grit before → after | Peak RSS Git, Grit before → after |
| --- | ---: | ---: | ---: | ---: | ---: |
| rust-lang/cargo | 3,078 | 10.5 | 23.5 → **17.3** | 3.2, 8.2 → **6.4** | 6, 15 → 15 MiB |
| git/git | 4,857 | 6.6 | 23.0 → **9.6** | 3.3, 18.7 → **5.2** | 6, 16 → 15 MiB |
| synthetic depth-7 tree | 48,750 | 54.9 | 59.7 → **48.2** | 18.0, 33.3 → **26.6** | 17, 23 → 23 MiB |
| 60 × 4 MB binaries, `core.filemode=false` | 60 | 1.3 | 358.1 → **2.4** | 1.3, 353.1 → **2.2** | 4, 20 → 13 MiB |
| 200 × 1 MB text, `core.filemode=false` | 200 | 1.3 | 377.0 → **2.6** | 1.3, 381.3 → **2.4** | 4, 14 → 13 MiB |

`grit-bench status` (synthetic 1k/10k/50k, committed as `status-clean-tree-before.json` → `status-clean-tree-after.json`): `status-clean-1000` 7.0 → 3.9 ms and `status-dirty-1000` 6.9 → 4.1 ms; the 10k rows no longer alternate with a ~33 ms racy re-hash mode; 50k rows are unchanged within noise.

```bash
./target/release/grit-bench --format json --output grit-utils/baselines/status-clean-tree-after.json status --sizes 1000,10000,50000
```

## Status right after a checkout (2026-10-25)

A clone, switch or `reset --hard` writes every file and then the index within the same second, so every entry is racily clean. Under Git's default whole-second rule, the next status has to re-hash all of them once, and then rewrite the index. Git does this too on Linux and macOS. Windows no longer pays it at all: see [On Windows: nanosecond timestamps](#on-windows-nanosecond-timestamps-2026-10-28) below. Dogfooding on Windows (issue #1914) measured that first `grit status` at 3–13 s on rust-lang/cargo, against 0.2 s for Git. Grit read the files one at a time, inside the serial pass over the stat scan, with a second `lstat` for each file, and on Windows every file open is expensive. Now `grit-lib` collects the racy entries from the scan, hashes them on the `IndexRefresh` worker pool, and applies the results from the scan's own stat data. The one-second rule is unchanged. A same-second edit with the same size is still caught by its content.

Same Linux VM sitting (4 vCPUs, ext4), release builds, system `git 2.43.0`. Each run starts from the index exactly as the clone left it. Times are hyperfine means in ms; "cold" drops the page cache before each run. Peak RSS comes from `/usr/bin/time`:

| Repository (tracked files, all racy) | Git | Grit before → after | Cold: Git, Grit before → after | Peak RSS Git, Grit before → after |
| --- | ---: | ---: | ---: | ---: |
| rust-lang/cargo (3,078) | 64.7 | 47.6 → **30.0** | 194.7, 178.3 → **106.6** | 8.4, 18.7 → 20.1 MiB |
| git/git (4,857) | 156.8 | 132.6 → **68.5** | 353.3, 325.4 → **164.2** | 7.7, 20.5 → 23.3 MiB |

After the first run, status on a clean tree is unchanged: 16.8 ms on cargo and 5.2 ms on git/git. On cargo, `grit switch --detach <old>` followed by `grit status` drops from 56 to 28 ms. The extra 1.5–3 MiB of peak memory belongs to the worker threads: with `index.threads=1`, peak RSS matches the old serial path. On Windows the parallel opens are where most of the gain comes from.

`grit-bench status` now includes a `status-after-checkout-{N}` row. Its hyperfine prepare step restores the racily clean index before each run. The committed results are `status-after-checkout-before.json` → `status-after-checkout-after.json`: `status-after-checkout-10000` goes from 50.2 to **27.9** ms (Git 31.6), and `status-after-checkout-1000` from 7.5 to **6.0** ms (Git 5.0). The dirty and clean rows are unchanged within noise.

```bash
./target/release/grit-bench --format json --output grit-utils/baselines/status-after-checkout-after.json status --sizes 1000,10000
```

### On Windows: nanosecond timestamps (2026-10-28)

The worker pool made the re-hash cheaper, but on Windows opening a file is the expensive part, so the first `grit status` after a clone still took 1.2–2.5 s on git/git and about 1 s on cargo (issue #2175), against 72 ms and 155 ms for Git for Windows, which stays stat-only. Waiting before that first run changes nothing: the comparison is between each file's mtime and the index file's, not the clock.

`core.usenanosec` now defaults to on for Windows (`StatMatchPolicy::DEFAULT_USE_NANOSEC`) and stays off elsewhere, like Git. NTFS keeps 100 ns timestamps, so an index written after the files is newer than every entry, and none of them is racy. Both racy detection and stat equality use the sub-second fractions, so a same-second, same-size edit still shows up through its mtime. Racy detection only uses them when stat equality does too: under `core.checkstat=minimal` both stay whole-second. Two smaller fixes go with it. Checkout's index-mtime bump now steps 100 ns on Windows, because NTFS truncated the old 1 ns step back onto the entry's own tick. The bump never moves into the next second: if the newest file sits on the last tick of its second, the index is left alone. And the Windows directory-scan shortcut takes one path `lstat` when only the sub-second fractions disagree, instead of re-hashing the file. Setting `core.usenanosec=false` brings back the whole-second rule.

Without a Windows runner, the table below simulates the Windows default on the same Linux VM with `core.usenanosec=true` (4 vCPUs, ext4, release builds, system `git 2.43.0`). Each run starts from the index exactly as `grit clone` (or `git clone`, for Git) left it. Times are hyperfine means in ms, and peak RSS comes from `/usr/bin/time`:

| Repository (tracked files) | Git | Grit, whole-second (Linux default) | Grit, `core.usenanosec=true` (Windows default) | Peak RSS Git, Grit whole-second → nanosecond |
| --- | ---: | ---: | ---: | ---: |
| rust-lang/cargo (3,078) | 62.7 | 30.3 | **18.6** | 7.8, 20.2 → 17.9 MiB |
| git/git (4,857) | 116.0 | 67.1 | **11.3** | 6.6, 25.1 → 18.1 MiB |

Under nanoseconds the first status on git/git opens 225 files (ignore and attribute files) instead of 5,093. On Linux, nothing changes by default: the whole-second column matches main within noise (69.5 → 67.1 ms on git/git, 30.8 → 30.3 ms on cargo), and so does a warm clean status (5.0 → 4.9 ms, Git 6.4).

## Adding many untracked directories (2026-10-25)

`grit add .` on a tree of thousands of untracked directories was quadratic in the directory count (issue #1854): 4,000 one-file directories took 90 s and 1 GB where `git add -A` takes 34 ms. Any pathspec, `.` included, made the untracked walk collapse each wholly untracked directory to `dir/` and then re-walk the whole work tree once per collapsed directory to list its files. Two changes in `grit-lib`:

- `add` treats `.` and `./` as no pathspec, and asks the walk for every untracked file in one pass. Collapsed `dir/` entries now only come from a normal-mode untracked cache on the index, which `add` reuses instead of rebuilding.
- Expanding collapsed directories (`expand_untracked_for_staging`) walks each one once, starting at the directory itself, and filters its files by the pathspecs. `stash -u` and the commit staging engine no longer expand at all, because their walks already list every file.

Same Linux VM sitting, release builds, system `git 2.43.0`, `/usr/bin/time` wall time and peak RSS, empty index (`git init` then add). Grit is `origin/main` before → after:

| Layout | Git | Grit before → after | Peak RSS Git, Grit before → after |
| --- | ---: | ---: | ---: |
| 1,000 one-file directories | 0.01 s | 4.97 → **0.02** s | 5, 79 → **16** MiB |
| 2,000 one-file directories | 0.02 s | 21.2 → **0.03** s | 5, 275 → **19** MiB |
| 4,000 one-file directories | 0.04 s | 90.5 → **0.06** s | 6, 1,059 → **23** MiB |
| 49,000 files in `d<i%200>/s<i%37>` | 0.29 s | 18.5 → **0.59** s | 14, 668 → **84** MiB |

With one tracked file committed first, `grit add 'd*'` on 2,000 directories went from 21.7 to 0.04 s, and `grit commit` (stage all) from 2.25 to 0.06 s. The staged index is byte-identical to `git add -A` (`git ls-files -s`). `grit-bench add-many-dirs` records the 4,000-directory and 49,000-file rows in `add-many-dirs.json`:

```bash
./target/release/grit-bench --format json --output grit-utils/baselines/add-many-dirs.json add-many-dirs --dirs 4000 --nested-files 49000
```

## Delete-heavy switch and hard reset (2026-10-24)

On Windows with real-time antivirus, deleting about 2000 files with `grit switch` or `grit reset --hard` took several seconds where Git took half a second (issue #1754). The cost was per-file work that opens files, which antivirus scanning makes expensive, rather than the deletes themselves:

- **`reset --hard` diffed the whole worktree for nothing.** Planning a hard reset ran a full index-to-worktree diff to list the edits it would discard, then threw the list away unless `--dry-run` was set. That diff stat'ed every tracked file and read every racily clean one, including the files about to be deleted. It now runs only for dry runs.
- **`reset --hard` stat'ed each file about three times.** Paths leaving the index were stat'ed and removed one by one, and every unchanged entry went through a per-path up-to-date check that stat'ed it, possibly hashed it, and stat'ed it again. Now tree-diff deletions go through the tree-diff checkout (whole directories removed at once), unchanged entries get one directory-batched stat pass, and only stat-dirty or racily clean entries are content-checked, in parallel. The scan only decides: rows written back to the index keep the stat data a path stat recorded, so directory-enumeration stats (which can differ from `lstat` on NTFS) never land in the index, and a directory that replaced a tracked file is never mistaken for it.
- **`switch` hashed racily clean files one at a time.** Right after a checkout, the files it wrote share the index's timestamp second, so switching away has to compare their content (Git does the same). The guard re-stat'ed each one and hashed them serially. It now reuses the directory scan's stat and hashes them on the index-refresh thread pool.

The hard-reset rework also fixed a correctness hole: raciness of unchanged entries was judged against the index written halfway through the reset, so a same-size edit made within the index's timestamp second could survive `reset --hard` and be recorded as clean.

Syscalls on the issue's synthetic tree (480 directories of 100 files, the target commit deletes 20 directories), run straight after the reverse checkout so the deleted files are racily clean, counted with `strace -f`:

| Command | Git | Grit before | Grit after |
| --- | ---: | ---: | ---: |
| `reset --hard`: stat calls | 46 133 | 146 883 | **46 341** |
| `reset --hard`: files opened | 0 | 2 000 | **0** |
| `switch`: stat calls | 54 650 | 6 342 | **4 347** |
| `switch`: racy files read | 2 000 | 2 000 (1 thread) | 2 000 (pool) |

`grit-bench daily` rows on the same shape (the fixture is packed, each run goes to the delete-heavy commit and back), same VM sitting, system `git 2.43.0`, committed in **`daily-delete-heavy-before.json`** and **`daily-delete-heavy-after.json`**:

| Scenario | Git ms (before / after run) | Grit ms before → after | Grit ÷ Git before → after | Grit peak RSS before → after |
| --- | ---: | ---: | ---: | ---: |
| `reset/hard-delete-heavy@wide-48k` | 305.3 / 306.0 | 557.3 → **276.0** | 1.83× → **0.90×** | 62 → **49** MB |
| `switch/delete-heavy@wide-48k` | 180.7 / 179.6 | 145.4 → 138.1 | 0.80× → 0.77× | 32 → 33 MB |

On Linux, opening a file is cheap, so the switch row only moves within noise. The win there is on platforms where each open is slow, and it can't be timed on this VM.

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench daily --only switch/delete-heavy@wide-48k,reset/hard-delete-heavy@wide-48k \
  --min-runs 10 --output grit-utils/baselines/daily-delete-heavy-after.json
```

## Reset reuses the index it replaces (2026-10-25)

The daily census had `reset/mixed@synth-10k` (`grit reset HEAD` on a tree with 1% of files edited) at 5× Git. A mixed reset built the new index from the target tree with empty stat data, so the refresh that followed opened and hashed all 10 000 tracked files (`strace`: 10 055 `openat`, 20 051 `read`, 30 358 `statx`) where Git only stats them.

- **Mixed reset seeds the new index from the current one.** Rows whose path, mode and blob id are unchanged keep their cached stat data, so the refresh hashes only files whose stat changed (156 `openat`, 251 `read`, one `statx` per file).
- **Unchanged subtrees come from the cache-tree, not the object database.** Like Git's one-way merge, flattening the target tree copies the current index's rows for every directory whose cache-tree node is valid and names the same tree, after checking that the node's entry count covers exactly the plain rows under that directory, so a stale node is never trusted. When the whole tree matches (`reset HEAD`), the rows are reused in place. Partial commits (`commit -- <paths>`) and `stash push` build their seeded index the same way.
- **`reset --hard` flattens the target tree the same way and keeps the cache-tree.** It used to read every tree object on every run and then write the index without a `TREE` extension, which Git primes after a hard reset. Without it, the next command had nothing to reuse.

Same VM sitting, release builds, system `git 2.43.0`, `origin/main` before → after (`grit-bench daily --min-runs 10`, committed in **`daily-reset-before.json`** and **`daily-reset-after.json`**):

| Scenario | Git ms (before / after run) | Grit ms before → after | Grit ÷ Git before → after | Grit peak RSS before → after |
| --- | ---: | ---: | ---: | ---: |
| `reset/mixed@synth-10k` | 8.8 / 9.1 | 45.6 → **12.2** | 5.17× → **1.35×** | 20.6 → **19.6** MiB |
| `reset/hard@synth-10k` | 16.4 / 16.5 | 21.8 → **20.1** | 1.33× → **1.22×** | 22.7 → 22.7 MiB |
| `reset/hard-delete-heavy@wide-48k` | 240.9 / 237.1 | 236.9 → 227.9 | 0.98× → 0.96× | 50.0 → 49.3 MiB |
| `commit/paths@synth-10k` | 22.8 / 22.7 | 28.6 → **25.5** | 1.25× → **1.12×** | 25.8 → 25.9 MiB |
| `stash/push-pop@synth-10k` | 213.7 / 212.7 | 158.6 → 148.4 | 0.74× → 0.70× | 26.7 → 26.3 MiB |

The mixed-reset samples spread from 10.2 to 14.5 ms in that run. Repeated with `hyperfine` on the dirty fixture without re-preparing, it is 43.6 → 9.9 ms (Git 7.9 ms). `status`, `switch`, `restore` and the other `commit` rows are unchanged within noise. The mixed reset's heap peak dropped from 3.9 to 2.8 MB (massif). Most of the remaining RSS gap to Git is the `grit` binary's mapped code pages (about 10.6 MB of the 16.7 MB resident at exit), not reset data.

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench daily \
  --only reset/mixed@synth-10k,reset/hard@synth-10k,reset/hard-delete-heavy@wide-48k,commit/paths@synth-10k,commit/small@synth-10k,commit/dry-run@synth-10k,stash/push-pop@synth-10k,status/dirty@synth-10k,restore/paths@synth-10k,switch/branch@synth-10k \
  --min-runs 10 --output grit-utils/baselines/daily-reset-after.json
```

## Content merges on both sides (2026-10-25)

A merge where both branches edited the same files grew with the number of content-merged paths times the size of the index: 146 s for 20 000 files edited on both sides, where Git takes 2.9 s, and 112 s for 2000 overlapping edits in a 200 000-file tree (issue #1881). For every path that needed a three-way content merge, the tree merge reloaded the index and rebuilt the attribute and ignore rules (config, `info/exclude`, the global ignore file, `.gitattributes`) up to four times, just to look up `merge=union`, `-merge`, `-text` and `-diff`. The rules are now built from one index load per tree merge, on the first path that needs them, and each path looks its attributes up once. Results are unchanged: the merged trees are identical to Git's.

Same VM sitting, release builds, system `git 2.43.0`, `/usr/bin/time` wall time and peak RSS, one run each:

| Repository | Git | Grit before | Grit after |
| --- | ---: | ---: | ---: |
| 2000 files, all edited on both sides | 0.36 s / 9 MB | 2.26 s / 25 MB | **0.24 s** / 25 MB |
| 20 000 files, all edited on both sides | 2.85 s / 40 MB | 145.9 s / 100 MB | **2.66 s** / 103 MB |
| 200 000 files, 2000 overlapping edits | 2.51 s / 96 MB | 112.2 s / 122 MB | **1.05 s** / 102 MB |

The new `grit-bench daily` row `merge/content-both-sides@wide-20k` (200 directories of 100 twenty-line files, `topic` appends a line to every file, `main` prepends one, packed fixture) is committed in **`daily-merge-content-wide-before.json`** and **`daily-merge-content-wide-after.json`**:

| Scenario | Git ms (before / after run) | Grit ms before → after | Grit ÷ Git before → after | Grit peak RSS before → after |
| --- | ---: | ---: | ---: | ---: |
| `merge/content-both-sides@wide-20k` | 1395.6 / 1378.3 | 149 610.9 → **1236.5** | 107.2× → **0.90×** | 115 → 116 MB |

Peak memory is still about twice Git's on this shape and is not addressed by this change.

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench daily --only merge/content-both-sides@wide-20k \
  --min-runs 10 --output grit-utils/baselines/daily-merge-content-wide-after.json
```

## Pick keeps the cache-tree (2026-10-24)

The daily census flagged `pick/one@synth-100k` going from 0.95× to 1.22× Git. The replay sequencer had started loading the index without its cache-tree extension to save memory. Without a cache-tree, the check that the index matches `HEAD` flattens the whole `HEAD` tree into a map, which costs more time and more memory than the extension it skipped. The index was also parsed a second time for the first replay step, and the index written after the pick lost its cache-tree. Pick and revert now keep the cache-tree, and the snapshot from the cleanliness check feeds the first step.

Same VM sitting, release builds, system `git 2.43.0`, `origin/main` before → after:

| Scenario | Git ms | Grit ms before → after | Grit peak RSS before → after |
| --- | ---: | ---: | ---: |
| `pick/one@synth-100k` (daily) | 208.1 | 245.8 → **180.5** | 63.5 → **47.6** MiB |
| `pick-100000` | 204.0 | 241.4 → **154.5** | 65.3 → **49.4** MiB |
| `revert-100000` | 195.9 | 241.5 → **155.9** | 66.1 → **51.1** MiB |
| `pick-series-100000` | 1014.5 | 3822.5 → **2057.9** | 67.0 → **44.2** MiB |
| `pick-10000` | 140.5 | 75.2 → **73.4** | 29.2 → **26.1** MiB |
| `revert-10000` | 136.1 | 81.0 → **72.2** | 29.6 → **26.9** MiB |
| `pick-series-10000` | 137.6 | 495.3 → **288.2** | 25.8 → **22.9** MiB |
| `pick-range-20` | 135.3 | 116.7 → **113.5** | 24.2 → **22.3** MiB |

`switch-*` and `merge-*` rows are unchanged within noise. `hot-paths-after.json` is refreshed from the after run.

## Pick, revert and merge apply path (2026-10-22)

A clean `grit pick` / `grit revert` of a commit touching 2000 paths spent most of its time on work that grows with the number of touched paths, not on I/O:

- **Sparse tree merge.** "Is anything under this directory touched?" and "does the other side have a file under this path?" scanned every touched path, once per tree entry visited and once per file in the directory/file pass. A fallback re-walked each side tree from the root, cloning parsed trees along the way. A per-merge prefix index makes these hash lookups. Parsed trees are now shared, not copied. The merged tree is identical.
- **Index after the merge checkout.** Checkout wrote the index, with fresh stat data for every file it wrote. Folding in the merge result then removed and re-inserted every touched row (one O(n) vector drain each) and re-stat'ed every file. A second index write re-hashed every same-second entry as racily clean. Rows checkout already left identical are now kept, and the index is written once, with checkout's racy policy. `post-index-change` runs once.
- **Clean-worktree guard.** The dense index-to-worktree pass lstat'ed every tracked file twice. It now trusts the parallel scan's stat for unchanged, non-racy entries.

`grit-bench hot-paths` (default sizes, warmup and runs), Grit before → after on the same VM sitting, vs system `git 2.43.0`. The refreshed numbers are committed in **`hot-paths-after.json`**:

| Scenario | Git ms | Grit ms before → after | Grit ÷ Git before → after |
| --- | ---: | ---: | ---: |
| `pick-10000` | 110.3 | 203.5 → **58.4** | 1.80× → **0.53×** |
| `revert-10000` | 105.7 | 239.2 → **60.2** | 2.25× → **0.57×** |
| `merge-10000` | 18.6 | 28.5 → **22.6** | 1.53× → **1.21×** |
| `pick-series-10000` | 111.0 | 512.4 → **300.7** | 4.60× → **2.71×** |
| `pick-100000` | 185.0 | 1802.3 → **186.6** | 9.33× → **1.01×** |
| `revert-100000` | 174.6 | 1765.7 → **185.0** | 9.37× → **1.06×** |
| `merge-100000` | 138.9 | 138.4 → **107.1** | 0.83× → **0.77×** |
| `pick-series-100000` | 868.2 | 4985.0 → **2451.9** | 5.78× → **2.82×** |

`switch-*` rows are unchanged within noise. Peak RSS on `pick-10000` stays around **26 MiB**, against Git's 9 MiB. About 10 MiB of that is resident `grit` binary text, and the heap peak is one transient copy of the index during checkout. Memory is not addressed by this change.

```bash
./target/release/grit-bench hot-paths --sizes 10000,100000 --format json --output grit-utils/baselines/hot-paths-after.json
```

## Tags on many refs (2026-10-24)

On the many-refs fixture (11 668 packed refs, 3 333 of them tags, half annotated with loose tag objects), `grit tag` spent almost all of its time classifying tags. Every tag ref cost an object-header read, and every annotated tag cost a second header read plus a full read of the tag object to peel it. That meant thousands of loose-object inflates, whose zlib setup alone was a fifth of the run.

- **Listing from peel lines.** `packed-refs` written with the `peeled` or `fully-peeled` trait already records each annotated tag's peeled object, which is how Git avoids those reads. [`list_tags_with_detail`](https://docs.rs/grit-lib/latest/grit_lib/repo_handles/struct.Tags.html#method.list) uses [`TagDetail::Names`](https://docs.rs/grit-lib/latest/grit_lib/porcelain/tag/enum.TagDetail.html#variant.Names) for the human names-only path and [`TagDetail::Kind`](https://docs.rs/grit-lib/latest/grit_lib/porcelain/tag/enum.TagDetail.html#variant.Kind) for classification from peel records. Loose overrides and files without a peel trait fall back to objects. [`TagDetail::Annotation`](https://docs.rs/grit-lib/latest/grit_lib/porcelain/tag/enum.TagDetail.html#variant.Annotation), used by `--json` and `--markdown`, still reads tag objects and preserves their documented fields.
- **Creating through the repository's ref store.** `create_tag` checked for an existing tag, and wrote the new one, through helpers that each opened a fresh ref store and so re-read and re-indexed `packed-refs`. The existence check, the write and the CLI's `HEAD` lookup now share the repository's store and its cached view. Sorted `packed-refs` files are binary-searched in place, without a line index.

`grit-bench daily` rows (20 runs per side, release build), current main before → after in one sitting against system `git 2.43.0`, committed in **`daily-tag-many-refs-before.json`** and **`daily-tag-many-refs-after.json`**:

| Scenario | Git ms | Grit ms before → after | Grit ÷ Git before → after |
| --- | ---: | ---: | ---: |
| `tag/list@many-refs` (human names) | 3.28 | 13.25 → **5.44** | 4.03× → **1.66×** |
| `tag/create@many-refs` | 2.21 | 4.36 → **4.35** | 1.97× → **1.97×** |
| `branch/create@many-refs` | 2.21 | 4.35 → **4.35** | 1.97× → **1.96×** |

The annotation-rich JSON path is not slower: hyperfine (20 runs) measured **11.6 → 10.6 ms**. Branch create is unchanged within noise, guarding the shared packed-ref path. Peak RSS for human tag listing fell from **21.1 to 15.3 MiB** in the daily rows; create remained within noise. The remaining gap to Git is mostly process start-up and CLI setup, tracked under the binary-diet follow-up in **Methodology** below.

```bash
cargo build --release -p grit-cli -p grit-utils
for id in tag/list@many-refs tag/create@many-refs branch/create@many-refs; do
  ./target/release/grit-bench daily --only "$id" --min-runs 20 --format json --output "/tmp/$(echo "$id" | tr '/@' '__').json"
done
```

## Incremental fetch negotiation (2026-10-25)

An incremental `grit fetch` or `grit pull` of a few new commits into a `git/git` clone took about 740 ms against 115 ms for Git. Nearly all of it was negotiation: before sending its first `have`, a protocol v2 fetch walked the client's entire history, inflating every commit from its pack, only to send the first 16 commits and find the server ready. Now [`Remote::fetch`](https://docs.rs/grit-lib/latest/grit_lib/remote/struct.Remote.html) walks history lazily, one round at a time. It sends 16 haves, then twice as many per round. Each round repeats the commits the server acknowledged, because a v2 server keeps no state between requests. It stops when the server is ready, the haves run out, or 256 haves go unacknowledged after a first acknowledgement. Negotiation also starts from remote-tracking refs, as Git's does, so commits already held under `refs/remotes/` are not downloaded again. For a `git/git` clone whose branch was 100 commits behind its remote-tracking ref, that cut the pack for 5 new commits from 105 commits to 5.

`grit-bench daily` rows (10 runs per side, release build), `origin/main` before → after in one sitting against system `git 2.43.0` on the factory VM (4 vCPU), committed in **`daily-fetch-pull-git-git-before.json`** and **`daily-fetch-pull-git-git-after.json`**. The `git/git` remote is served by `git http-backend`, and `fetch/incremental-http` by `grit-http-server`:

| Scenario | Git ms | Grit ms before → after | Grit ÷ Git before → after | Grit peak RSS before → after |
| --- | ---: | ---: | ---: | ---: |
| `pull/check@git-git` | 116.5 | 739.9 → **52.9** | 6.25× → **0.45×** | 67.3 → **50.1 MB** |
| `pull/fast-forward@git-git` | 129.6 | 729.9 → **57.6** | 5.94× → **0.44×** | 66.6 → **50.1 MB** |
| `fetch/incremental@git-git` | 114.9 | 729.6 → **52.0** | 6.29× → **0.45×** | 66.6 → **49.9 MB** |
| `fetch/one-branch@git-git` | 116.3 | 740.3 → **50.8** | 6.13× → **0.44×** | 66.7 → **49.8 MB** |
| `fetch/dry-run@git-git` | 113.0 | 732.2 → **49.5** | 6.25× → **0.44×** | 67.3 → **49.7 MB** |
| `fetch/incremental-http@git-git` | 16 069 | 15 583 → **13 513** | 0.97× → **0.84×** | 66.9 → **67.2 MB** |
| `fetch/noop@git-git` | 25.7 | 11.0 → **12.1** | 0.45× → **0.47×** | 17.3 → **17.8 MB** |
| `fetch/deepen@git-git` | 18 070 | 16 229 → **16 182** | 0.86× → **0.90×** | 222.7 → **228.7 MB** |
| `pull/rebase@synth-10k` | 76.4 | 27.5 → **27.4** | 0.36× → **0.36×** | 43.6 → **44.5 MB** |
| `clone/http@synth-10k` | 856.1 | 743.9 → **746.5** | 0.91× → **0.87×** | 30.3 → **31.3 MB** |

The rows that did not negotiate a deep local history are unchanged within noise: a 30-run re-measurement of `fetch/noop` gave 11.99 → 11.95 ms. The `fetch_v2_negotiation_rounds` integration test runs against `git daemon`. It fails if a fetch the server can answer after one round reads deep local history, if a later round drops the acknowledged commits, or if commits already held under `refs/remotes/` are sent again.

```bash
cargo build --release -p grit-cli -p grit-utils -p grit-http-server
./target/release/grit-bench daily --only pull/check@git-git,pull/fast-forward@git-git,fetch/incremental@git-git \
  --min-runs 10 --http-server ./target/release/grit-http-server --format json --output /tmp/fetch-pull.json
cargo test -p grit-lib --test integration fetch_v2_negotiation_rounds::
```

## Serving an incremental fetch: thin packs (2026-10-27, #2135)

Answering a fetch of a few new commits from `git/git` took seconds in `grit upload-pack`, longer with a pack bitmap than without, and the thin pack was up to 7× bigger than Git's ([#2135](https://github.com/gitbutlerapp/grit/issues/2135)). Thin bases were only ever searched for blobs, and with a bitmap the search scanned the client's whole object set for every changed blob. Changed trees went out whole.

Pack generation now picks thin-pack bases the way Git's preferred bases work. The edge commits are the client's commits that the new history builds on. The trees and blobs at the changed paths in their trees join the delta search as bases that are never written. A delta against one goes out as `REF_DELTA`, so a changed tree costs tens of bytes instead of its full size. Only the client's directories on the way to a changed path are read, matched by pack name hash. Further changes:

- **Stored deltas.** On-disk deltas whose base the client holds are still reused. When the client also holds an older copy at the same path, the object goes through the search instead. An object that is the base of a reused delta may become a delta against a client-held base, with room left for its dependent chain.
- **Commit walk.** Upload-pack stops walking the client's history where it meets the wants, as push already did, instead of reading every ancestor of every `have`.
- **Bitmaps.** The bitmap walk for tips without a stored bitmap goes newest first and stops at objects a stored bitmap already covers. The want side starts from the have side's objects, so the have closure is computed once. The bitmap uses the shared pack index instead of re-reading and re-hashing the `.idx`.

Factory VM, 4 vCPU, Linux 6.12, Rust 1.99, `release` profile, system Git 2.43.0. The fixture is a `git clone --bare` of upstream `git/git` (82,605 commits, 422k objects) after `git repack -adb` and `git commit-graph write --reachable`, and a copy with the `.bitmap` removed. One protocol v2 `command=fetch` request (want `HEAD`, have `HEAD~N`, `thin-pack`, `ofs-delta`, `done`) goes to `--stateless-rpc` of each server. The table shows the median of three runs and the largest `wait4` peak RSS. Every Grit pack passes `git index-pack --fix-thin`. Main is `9a9e069ca`.

| Request | System Git | Grit main | Grit after |
| --- | ---: | ---: | ---: |
| `HEAD~1`, bitmap | 11 ms / 23 MB / 28.9 KB | 3,015 ms / 147 MB / 38.3 KB | 26 ms / 48 MB / 1.7 KB |
| `HEAD~1`, no bitmap | 7 ms / 13 MB / 1.7 KB | 616 ms / 79 MB / 38.3 KB | 17 ms / 41 MB / 1.7 KB |
| `HEAD~20`, bitmap | 23 ms / 37 MB / 536 KB | 4,943 ms / 200 MB / 1,083 KB | 114 ms / 91 MB / 145 KB |
| `HEAD~20`, no bitmap | 62 ms / 52 MB / 145 KB | 757 ms / 99 MB / 1,018 KB | 92 ms / 83 MB / 145 KB |
| `HEAD~100`, bitmap | 38 ms / 55 MB / 4.87 MB | 12,426 ms / 373 MB / 6.23 MB | 272 ms / 188 MB / 1.92 MB |
| `HEAD~100`, no bitmap | 369 ms / 130 MB / 2.07 MB | 1,328 ms / 202 MB / 5.85 MB | 314 ms / 195 MB / 2.00 MB |
| `HEAD~300`, bitmap | 57 ms / 74 MB / 10.80 MB | 26,992 ms / 454 MB / 13.53 MB | 458 ms / 225 MB / 4.52 MB |
| `HEAD~300`, no bitmap | 782 ms / 205 MB / 4.43 MB | 1,467 ms / 270 MB / 12.93 MB | 677 ms / 273 MB / 4.56 MB |

Without a bitmap, Grit now sends the same pack size as Git, at about Git's speed for larger ranges. With a bitmap, Git copies stored deltas and skips the delta search, so it answers faster but sends packs up to 2.4× bigger. Grit still runs the search, which costs time and memory but sends about the same pack as without a bitmap. A full clone is unchanged: the packs are byte-identical, and the bitmap clone took 4.4 s against 5.6 s on main (Git 2.4 s).

The `thin_upload_pack_*` tests in `grit-lib/tests/pack_objects_git_compat.rs` build a wide-tree history and check the result with and without a bitmap. The client's root tree, changed subtree and the old version of a changed file must be `REF_DELTA` bases. A one-commit pack may hold no more whole objects than `git pack-objects --thin`. Every pack must stay within 1.25× of Git's size, pass `git index-pack --fix-thin` and `git fsck --strict`, and the bitmap pack may not be bigger than the one without.

```bash
cargo test -p grit-lib --features full-tests --test integration thin_upload_pack
```

## Receiving a pack: clone and fetch memory (2026-10-25)

Cloning over HTTPS peaked at several times the received pack's size ([#1672](https://github.com/gitbutlerapp/grit/issues/1672)). The received pack is spooled to disk and indexed through a memory map, and peak RSS counts the map's resident pages. The trailer checksum, the object scan and delta resolution each read the whole map, so the full pack stayed resident while it was indexed. After indexing, a check for thin-pack bases re-inflated every object again on one thread. Completing a thin pack then read the pack into memory and indexed it a second time.

Indexing now keeps the pack's resident pages within a few megabytes. The two sequential passes drop the map's pages every 2 MiB, and delta resolution reads each compressed delta from the file with a positioned read, as `git index-pack` does. The index build reports the thin-pack bases it took from the object database. Missing bases are appended to the pack file in place, the header count and checksum are rewritten, and records for the new objects are added to the index without rebuilding it. Smaller buffers also cut working memory: deltas waiting to be resolved take 32 bytes instead of 72, worker threads hand their records over in batches, and the `.idx` is sorted in place and streamed to disk. The thin-pack rewrite also fixes a compatibility bug. Before it, the `.idx` kept the old pack checksum, and `git fsck` reported "packfile does not match index" after an incremental `grit fetch` from GitHub.

Factory VM (4 vCPU), system `git 2.43.0`, release builds, `origin/main` (`2657344c7`) against this change in one sitting, median of three runs. Clone rows are `grit clone` / `git clone -q` from GitHub. Ingest rows index the pack that `git clone` received: [`install_pack_path`](https://docs.rs/grit-lib/latest/grit_lib/objects/index_pack/fn.install_pack_path.html) with fetch options against `git index-pack` on the same file.

| Scenario | Git | Grit before | Grit after |
| --- | ---: | ---: | ---: |
| clone `rust-lang/cargo` (78 MB pack) | 4.54 s / 46.6 MB | 3.42 s / 126.9 MB | **2.40 s / 72.4 MB** |
| clone `git/git` (326 MB pack) | 18.38 s / 96.8 MB | 10.69 s / 419.2 MB | **7.47 s / 167.4 MB** |
| ingest `rust-lang/cargo` pack | 3.06 s / 28.5 MB | 2.39 s / 117.3 MB | **1.28 s / 35.1 MB** |
| ingest `git/git` pack | 13.26 s / 63.5 MB | 8.75 s / 406.6 MB | **5.21 s / 77.4 MB** |

Ingest now peaks at 1.2× `git index-pack`, down from 4–6×, and runs in 40% less time. The rest of the clone peak comes from checkout, which reads blobs through the 96 MB delta-base cache (`core.deltaBaseCacheLimit`, the same default as Git).

Two gates keep the memory from regressing. The `pack_ingest_memory` integration test builds a 43 MB, 70k-object pack with `git fast-import` and fails when grit's ingest peaks above 1.5× `git index-pack` on it, each measured in its own child process. `origin/main` measures 5.7× there and this change measures 1.4×. The daily matrix row `clone/full@git-git` records time and peak RSS for a full `git/git` clone over smart HTTP from `git http-backend`. The `grit-bench network` clone, fetch, push and ls-remote scenarios now record peak RSS for both tools as well.

```bash
cargo test -p grit-lib --test pack_ingest_memory -- --nocapture
GRIT_CARGO_PACK=/path/to/received.pack GRIT_SKIP_VERIFY=1 \
  cargo test --release -p grit-lib --test integration pack_ingest_roundtrip::rss_harness_install_pack_path -- --ignored --exact
```

## Large files in the work tree (2026-10-25)

Writing one 350 MB file to the work tree took grit 8–9 s and about 1 GB of memory on `restore`, `switch` and `clone`, where Git took about a second and 93 MB ([#1849](https://github.com/gitbutlerapp/grit/issues/1849)). Three costs added up:

- **Loose inflate was quadratic.** The loose-object reader handed zlib the whole spare capacity of its output buffer on every call. The zlib binding zero-fills that space each time, and with one 64 KiB input chunk per call that meant re-zeroing hundreds of megabytes per chunk. The buffer is now grown and zeroed only ahead of the inflater, and it is sized straight to the size the object header declares.
- **The blob was copied three times.** The loose store copied the payload out of the inflate buffer. The work-tree conversion copied it again even when it changed nothing. Restore then re-read and copied the written file to check it.
- **Nothing streamed.** `Odb::open_stream` read every stream to the end before returning it.

Checkout now streams a regular file larger than 1 MiB straight from the object database into the work tree when no conversion applies to its path. That means no `ident`, `working-tree-encoding`, smudge filter or LF-to-CRLF setting; see `worktree_conversion_is_identity`. Loose objects and objects stored whole in a pack are inflated as the file is written ([`Odb::with_stream`](https://docs.rs/grit-lib/latest/grit_lib/objects/odb/struct.Odb.html#method.with_stream)). Pack deltas are still rebuilt in memory, as in Git. Files at or under 1 MiB, which is nearly all of them, are still read with one lookup: a loose object's header, or a whole pack entry's header, gives the size before the body is inflated, so the size check is part of the read. A file that does stream is looked up twice, once to size it and once to open the stream, and a whole pack entry's bytes are checked against the pack index CRC before streaming, so a corrupt copy is never streamed. Streamed data goes to a temporary file next to the destination, which is renamed into place only after the object's size and zlib checksum have been verified. A corrupt object fails the checkout and leaves the old file, or no file, in place. Hashing a work-tree file (the racy-clean checks in `switch`, `restore` and `status`) no longer copies files whose content goes into Git unchanged, and hashes files of 1 MiB or more in chunks.

Factory VM (4 vCPU), system `git 2.43.0`, release builds, `origin/main` (`5d5136859`) against this change in one sitting, `/usr/bin/time` wall time and peak RSS. The repository is the issue's: one 349 MB text file (`seq 1 40000000`) committed loose.

| Command | Git | Grit before | Grit after |
| --- | ---: | ---: | ---: |
| `restore big.txt` | 0.95 s / 91 MB | 9.33 s / 1011 MB | **0.78 s / 14 MB** |
| `switch main` (writes the file) | 0.94 s / 91 MB | 8.21 s / 1012 MB | **0.48 s / 15 MB** |
| `restore` of 300 MB random bytes, packed | 0.43 s / 605 MB | 1.53 s / 1213 MB | **0.45 s / 314 MB** |
| `clone` of a packed repository with both files | 1.59 s / 689 MB | 2.12 s / 1963 MB | **0.72 s / 399 MB** |

In the packed rows peak RSS includes the pages of the mapped pack read through while inflating, and the time includes checking the pack entry against its index CRC before it is streamed. Git reads a packed blob under `core.bigFileThreshold` (512 MiB) into memory whole.

The daily matrix row `switch/large-blob@big-asset` switches away from a 300 MB loose text file and back, so each run checks the racily clean file before deleting it and then writes it again. Committed in **`daily-large-blob-before.json`** and **`daily-large-blob-after.json`**:

| Scenario | Git ms (before / after run) | Grit ms before → after | Grit ÷ Git before → after | Grit peak RSS before → after |
| --- | ---: | ---: | ---: | ---: |
| `switch/large-blob@big-asset` | 870 / 905 | 4691 → **615** | 5.39× → **0.68×** | 915 → **16** MiB |

Checkouts of many small files did not change: switching both ways between two commits that differ in every file of a 30 000-file tree measured 1.195 s → 1.188 s packed and 1.167 s → 1.107 s loose (hyperfine, 20 runs), and cloning that repository 564 → 543 ms.

Local clones now link loose objects, or copy their compressed files when linking is unavailable, before considering a new pack. Their transfer-statistics walk reads large blobs' headers instead of inflating their bodies: blobs have no outgoing links to traverse. This removes the remaining allocation proportional to a loose blob's size during clone. Human `grit cat` output also streams blobs of at least 1 MiB through the library's replace-aware display API; JSON and Markdown retain their content fields.

Follow-up measurements on 2026-10-26 used an AMD EPYC 7763 VM (4 vCPU), system Git 2.55.0, and identical `dist` profiles for current main (`a1f4e2168`) and this change (`c858ca944`). Builds and tests finished before timing began. The fixture is the issue's `seq 1 40000000` (348,888,897 bytes), stored loose. Each large scenario ran three times; the small-file clone ran seven times. Times are medians and RSS is the maximum across runs, measured with `/usr/bin/time -v`. Clone results pass system `git fsck` and have a clean `git status`; restored and switched files pass `git diff --exit-code`.

| Command | Git seconds / MiB | Grit main seconds / MiB | Grit after seconds / MiB |
| --- | ---: | ---: | ---: |
| `restore big.txt` | 0.88 / 90.8 | 0.77 / 11.2 | **0.76 / 11.2** |
| `switch main` (writes the file) | 0.88 / 91.0 | 0.50 / 12.1 | **0.50 / 12.4** |
| Local `clone` | 0.88 / 91.2 | 1.01 / 676.2 | **0.50 / 13.4** |
| `cat HEAD:big.txt` (Git: `cat-file blob`) | 0.74 / 90.5 | 0.70 / 1340.1 | **0.54 / 9.9** |
| Local clone, 5,000 small files | 0.37 / 6.1 | 0.43 / 18.0 | **0.42 / 18.0** |

Raw samples, including current main in each scenario's `baseline_grit` field, are committed in `grit-utils/baselines/large-blob-materialization.json`. The `clone_large_blob_rss` integration test clones a 64 MiB loose blob through the public library API, checks that it is linked, runs system Git's fsck/status checks, and limits peak-RSS growth to 16 MiB. Before the statistics-walk fix it grew by 128 MiB and failed; afterward it passes.

Issue [#2379](https://github.com/gitbutlerapp/grit/issues/2379): streaming inflate must stay fast as well as lean. On a 400 MiB incompressible blob, pre-fix grit restore took about 2.0 s loose / 1.3 s packed (≈3–4× system Git) while holding ≈9 MiB RSS. The fix caps each zlib `decompress` input/output window (256 KiB), verifies pack index CRC32 while inflating (instead of a separate pass that dropped the whole mapping), advises `DontNeed` only on the pack prefix already consumed, streams commits from 1 MiB with zlib level 0 when the first chunk does not shrink, and keeps `checkout_large_blob_rss` green. The `checkout_incompressible_blob_restore` integration test restores a 96 MiB incompressible blob loose and packed and fails if grit is more than 2.75× Git wall time on the host.

The `checkout_large_blob_rss` integration test restores and switches to a 64 MiB blob, loose and packed, and fails if the process's peak RSS grows by 16 MiB or more. On the original pre-streaming baseline (`5d5136859`) it grows by 139 MiB. Packed-object RSS also includes resident pages of the compressed pack mapping; this can make the strict RSS-growth assertion sensitive to the process's earlier high-water mark.

```bash
cargo test -p grit-lib --test integration checkout_large_blob_rss:: checkout_large_blob_streaming::
./target/release/grit-bench daily --only 'switch/large-blob@big-asset' \
  --format json --output grit-utils/baselines/daily-large-blob-after.json --grit ./target/release/grit
```

## Performance plan close-out (2026-10-21)

After the memory and libification wave (fair single-process `grit log --limit`, mapped pack index / commit-graph / MIDX, packed-refs seek, repository path cache, LTO release profile, merge tree cache, pick/revert/merge pass reduction, status/diff scan memory, add without full index copies), the factory re-ran the **cooking dashboard** matrix on **`grit 0.5.15`** (`756aa953`) vs system **`git 2.43.0`**. Committed baselines: **`suite-after-LH.json`**, **`hot-paths-after.json`**, **`commit-after-LH.json`**, and **`dashboard/current.json`**. **`grit-bench compare`** against the pre-plan dashboard snapshot at **`grit 0.5.14`** shows **no overlapping 10k scenario regressed more than 5%** on Grit median wall time.

### Synthetic L/H dashboard (before → after)

| Scenario | Git ms (before→after) | Grit ms (before→after) | Ratio before→after |
| --- | ---: | ---: | ---: |
| `status-dirty-10000` | 11.8→10.3 | 13.8→8.7 | 1.17×→**0.84×** |
| `status-clean-10000` | 8.1→8.0 | 11.4→7.6 | 1.40×→**0.95×** |
| `add-10000` | 20.1→23.1 | 33.8→30.2 | 1.68×→**1.31×** |
| `commit-10000` | 199.7→176.4 | 113.9→73.1 | 0.57×→**0.41×** |
| `switch-10000` | 23.6→21.9 | 30.2→27.1 | 1.28×→**1.24×** |
| `pick-10000` | 123.3→113.5 | 671.5→199.7 | 5.45×→**1.76×** |
| `merge-10000` | 105.1→18.5 | 732.8→28.5 | 6.97×→**1.54×** |
| `diff-10000` | 5.7→5.3 | 12.4→10.3 | 2.19×→**1.92×** |
| `log-10000` | 3.3→3.2 | 55.9→4.4 | 17.0×→**1.39×** |

### Synthetic 10k — wall time and peak RSS (before → after)

Each row is **Git → Grit** on median wall time (ms) and **peak RSS** from hyperfine + `grit-bench measure-rss` (same capture as **`dashboard/current.json`**). On this VM, Grit peak RSS exceeds **1.5×** Git on **`status-*`**, **`pick-*`**, and **`log-10000`**; **`add`**, **`commit`**, **`switch`**, **`merge`**, and **`diff`** stay below **1.5×** on RSS while several remain above **1.5×** on wall time.

| Scenario | Git ms | Grit ms | Git RSS | Grit RSS | Ratio |
| --- | ---: | ---: | ---: | ---: | ---: |
| `status-dirty-10000` | 11.8→10.3 | 13.8→8.7 | 9.8 MiB | 15.6 MiB | 0.84× |
| `status-clean-10000` | 8.1→8.0 | 11.4→7.6 | 9.2 MiB | 15.5 MiB | 0.95× |
| `add-10000` | 20.1→23.1 | 33.8→30.2 | 16.4 MiB | 19.8 MiB | 1.31× |
| `commit-10000` | 199.7→176.4 | 113.9→73.1 | 17.1 MiB | 21.0 MiB | 0.41× |
| `switch-10000` | 23.6→21.9 | 30.2→27.1 | 14.6 MiB | 19.8 MiB | 1.24× |
| `pick-10000` | 123.3→113.5 | 671.5→199.7 | 16.1 MiB | 25.3 MiB | 1.76× |
| `merge-10000` | 105.1→18.5 | 732.8→28.5 | 15.3 MiB | 19.8 MiB | 1.54× |
| `diff-10000` | 5.7→5.3 | 12.4→10.3 | 15.6 MiB | 17.1 MiB | 1.92× |
| `log-10000` | 3.3→3.2 | 55.9→4.4 | 7.0 MiB | 14.7 MiB | 1.39× |

At **100k** files the same suites land in **`dashboard/current.json`** (22 scenarios). Highlights vs the previous **`suite-after-LH.json`** / **`hot-paths-after.json`** captures: dirty/clean **status** at 100k is near parity (**~0.98–1.26×**); **`switch-100000`** and **`merge-100000`** drop to **~0.7–1.0×**; **`switch-wide-100000`** falls from **~13×** to **~1.03×**. Remaining headline gaps above **1.0×** Git time on this refresh:

| Scenario | Grit ÷ Git | Profile / why |
| --- | ---: | --- |
| `revert-100000` | **9.96×** | Merge-engine revert still replays large tree diffs; pick/revert pass sharing helped 10k more than 100k topic commits. |
| `pick-100000` | **9.20×** | Same merge/checkout path as revert on ~2000-path topics. |
| `pick-series-100000` | **5.87×** | By design **20×** separate `grit pick` processes vs one `git cherry-pick` range (process + repo open overhead). |
| `pick-series-10000` | **4.31×** | Same multi-process model at L tier. |
| `diff-100000` | **2.87×** | Parallel worktree diff still allocates per-entry results on wide modified sets (~2000 paths); RSS not yet the limiter. |
| `diff-10000` | **1.96×** | Same diff pipeline at L. |
| `add-100000` | **1.86×** | Staging ~20% of 100k paths still walks the full index; incremental index updates reduced copies but not scan volume. |
| `pick-10000` / `revert-10000` | **1.73× / 1.88×** | Residual merge + index write cost on 2000-path topics. |
| `merge-10000` | **1.51×** | Small-repo merge still above Git on median; 100k merge is parity. |
| `add-10000` | **1.49×** | Index update path at L. |
| `log-10000` / `log-100000` | **~1.37×** | Single `grit log --limit 100` vs `git log -100 --oneline`; remaining gap is revwalk + formatting, not paging. |
| `status-clean-100000` | **1.26×** | Clean 100k worktree: untracked/ignore setup vs Git preload. |

Factory follow-ups (queued as **Speed:** tasks): `diff-100000`, `add-100000`, `pick-series-100000`, `pick-100000`, `revert-100000`, `pick-series-10000`, `add-10000`, `diff-10000`, `pick-10000`, `revert-10000`, `merge-10000`, `log-100000`. The `pick-*` and `revert-*` rows were closed on 2026-10-22. See [Pick, revert and merge apply path](#pick-revert-and-merge-apply-path-2026-10-22).

### Daily common paths (roadmap 2)

**`grit-utils/baselines/daily-common-paths.json`** records six **`grit-bench daily`** matrix ids added for roadmap-2 CLI conventions: **`diff/stat@synth-100k`**, **`log/stat-n50@git-git`**, **`log/json-filter@git-git`**, **`notes/list-n100@synth-10k`**, and shallow clone over **`file://`** and grit smart HTTP (**`clone/shallow-depth1@synth-10k`**, **`clone/shallow-depth1-http@synth-10k`**). Each row includes hyperfine medians and **`peak_rss_bytes`** for grit and system git; pre-timed hooks check output equivalence where semantics match (full diff numstat paths, per-commit log stat blocks, log JSON commit oids, exact note `(commit, blob)` mappings, shallow boundary refs and `.git/shallow`). On the factory VM capture (2026-10-23), **`log/stat-n50@git-git`** is the largest time gap (**~4.3×** git); that row is queued for the speed agent rather than fixed in the registry change.

```bash
cargo build --profile dist -p grit-cli -p grit-http-server -p grit-utils
./target/dist/grit-bench daily --only 'diff/stat@synth-100k,log/stat-n50@git-git,log/json-filter@git-git,notes/list-n100@synth-10k,clone/shallow-depth1@synth-10k,clone/shallow-depth1-http@synth-10k' \
  --format json --output grit-utils/baselines/daily-common-paths.json --grit ./target/dist/grit
```

### Clone checkout of a wide tree (2026-10-24)

The daily matrix id **`clone/checkout@synth-100k`** clones a bare copy of the **100k-file** synthetic fixture by plain path, so both tools link the pack locally and the timed run is dominated by checking out every file (`git clone -q <path>` vs `grit clone <path>`). It closes issue #1678, where a 200k-file clone was **2.5×** slower than git with **~4×** git's user CPU: the populated-tree checkout writes files in chunks of 2048, and every chunk rebuilt the index cache tree from the whole commit tree (about 100 full tree walks for 200k files) only to discard it until the last chunk. Chunks now update only the in-memory index; the cache tree is built and the index is written once after the last chunk.

| Capture | git | grit | grit ÷ git | grit peak RSS |
| --- | ---: | ---: | ---: | ---: |
| Before (`daily-clone-checkout-synth-100k-before.json`) | 2957 ms | 4269 ms | 1.44× | 150 MiB |
| After (`daily-clone-checkout-synth-100k-after.json`) | 2959 ms | 3155 ms | 1.07× | 150 MiB |

On the 200k-file repro from the issue (one commit, gc'd, 4 vCPU), `grit clone` went from **7.16 s** (user 5.34 s) to **3.12 s** (user 1.45 s), against **3.50 s** (user 1.14 s) for git; peak RSS is about **89 MB** vs **61 MB** for git. The remaining system time is concurrent file creation in the same directories across checkout workers.

```bash
./target/release/grit-bench daily --only 'clone/checkout@synth-100k' \
  --format json --output grit-utils/baselines/daily-clone-checkout-synth-100k-after.json --grit ./target/release/grit
```

### Log pages on git.git (2026-10-25)

`grit log` pages walk history in committer-date order, and the walk reads every commit it queues to learn its parents and date. The rows were then built by reading and parsing the same commits a second time, which cost about a third of a 5000-commit page. The walk now keeps the payloads it read (in one shared buffer, released as soon as the rows are loaded) and hands them to the row loader through [`LogWalkPage::commit_data`](https://docs.rs/grit-lib/latest/grit_lib/revwalk/log_walk/struct.LogWalkPage.html), so each commit is inflated once. Several fixed costs paid on every page also shrank: peeling to a commit decides the object kind from its header instead of inflating it twice, DWIM ref resolution reads each candidate (`refs/HEAD`, `refs/tags/HEAD`, …) once instead of twice, `packed-refs` records decode their ids while validating them, and pack index lookups in walk order binary-search directly instead of first comparing up to 128 rows after the previous hit.

Hyperfine medians (interleaved runs) of `dist` builds of `origin/main` (`fce0b9c5f`) and of this change rebased onto it, on the pinned git.git clone (factory VM, same sitting), with median peak RSS over 15 runs. Sub-5 ms rows shift by a few tenths of a millisecond with machine load, so compare the two grit columns, which were timed together:

| Command | git ms | grit main | grit after | git RSS | grit RSS main | grit RSS after |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| `log --limit 100 --json` (`log/json-100@git-git`) | 2.95 | 4.44 | 4.33 | 12.3 MiB | 18.8 MiB | 18.7 MiB |
| `log --limit 100` (`log/oneline-100@git-git`) | 3.03 | 3.97 | 3.91 | 12.2 MiB | 18.2 MiB | 18.2 MiB |
| `log --limit 50 --json --filter …` (`log/json-filter@git-git`) | 2.23 | 4.57 | 4.42 | 9.9 MiB | 17.2 MiB | 17.3 MiB |
| `log --limit 1000 --json` | 11.98 | 22.07 | 16.34 | 15.0 MiB | 23.5 MiB | 23.6 MiB |
| `log --limit 5000` (`log/all@git-git`) | 47.79 | 106.00 | 64.28 | 18.0 MiB | 31.4 MiB | 31.4 MiB |

On the headline row, five interleaved 200-run rounds put the change 1.0–5.5% ahead of `main` every time (median 4.43–4.49 ms → 4.23–4.37 ms), and `perf stat` task-clock over 200 runs drops from 3.84–3.87 ms to 3.69–3.78 ms. Path-limited, `--author`, `--stat`, `revs` and `show` rows are unchanged within noise in interleaved runs (`daily-log-git-git-before.json` / `daily-log-git-git-after.json` hold the full daily matrix capture for both binaries). What remains of the 100-commit gap is mostly fixed cost: argument parsing, repository discovery and config loading, and opening the pack for the first object.

The daily runner used to read the clock only when a 1 ms `try_wait` poll first saw the child gone, so every sample was rounded up to a whole poll tick (about 1.1 ms here): sub-5 ms rows could only report values like 3.29, 4.39 or 5.44 ms, and a few-tenths gain stayed invisible unless it crossed a tick. On Linux it now blocks in `waitid` and reads the clock when the child exits, while a watchdog thread keeps the RSS sampling and the timeout and memory caps. Sub-5 ms rows captured before that fix (for example older `daily/` files) carry that rounding. Pass `--baseline-grit` to time a second grit binary in the same run for paired headline comparisons:

```bash
cargo build --profile dist -p grit-cli -p grit-utils
./target/dist/grit-bench daily --only 'log/json-100@git-git,log/oneline-100@git-git' \
  --grit ./target/dist/grit --baseline-grit /path/to/main/grit --min-runs 20 --format json --output /tmp/log-paired.json
./target/dist/grit-bench daily --only 'log/json-100@git-git,log/oneline-100@git-git,log/all@git-git,log/author@git-git,log/json-filter@git-git,log/limit-path@git-git,log/path@git-git,log/stat-n50@git-git' \
  --min-runs 30 --format json --output grit-utils/baselines/daily-log-git-git-after.json --grit ./target/dist/grit
```

### git.git headlines (hyperfine wall time + peak RSS)

**`grit-utils/baselines/dashboard/git-dot-git-headlines.json`** uses the standard grit-bench schema (hyperfine medians + **`peak_rss_bytes`** per side). Headline wall times are **not** taken from the daily RSS poll loop (which quantized sub-50 ms commands to ~50 ms); they use the same hyperfine + **`measure-rss`** path as other scenario suites.

| Scenario | Git ms | Grit ms | Git RSS | Grit RSS | Ratio |
| --- | ---: | ---: | ---: | ---: | ---: |
| `log/oneline-100@git-git` | 3.44 | 5.36 | 14.0 MiB | 24.4 MiB | 1.56× |
| `show/head@git-git` | 2.02 | 3.77 | 7.4 MiB | 14.8 MiB | 1.87× |
| `resolve/head-n@git-git` | 1.25 | 3.21 | 7.4 MiB | 12.5 MiB | 2.57× |
| `branch/list@many-refs` | 1.41 | 8.61 | 7.2 MiB | 15.8 MiB | 6.10× |

**`dashboard/git-dot-git-worktree-status.json`** captures clean/dirty **status** (and related worktree rows) on the pinned **git.git** clone; every git/grit side includes **`peak_rss_bytes`** (e.g. clean status **~27 MiB** on both tools on this VM).

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench fixtures prepare --only git-git,git-git-wt,many-refs
./target/release/grit-bench status --sizes 10000,100000 --format json --output /tmp/status.json
./target/release/grit-bench hot-paths --sizes 10000,100000 --format json --output grit-utils/baselines/hot-paths-after.json
./target/release/grit-bench headlines --offline --format json --output grit-utils/baselines/dashboard/git-dot-git-headlines.json
```

## grit-bench vs Git

### Methodology

- **Build profile:** Command-level numbers use a **`dist`** **`grit`** (**`cargo build --profile dist -p grit-cli`**, or **`make build`**): **`lto = "fat"`**, **`codegen-units = 1`**, **`opt-level = 3`**, **`strip = true`**, **`panic = "unwind"`** (required because **`grit-lib`**'s parallel pool uses **`catch_unwind`** / **`resume_unwind`**). Workspace **`[profile.release]`** stays fast for **`cargo test --release`** (no fat LTO on test binaries). Rebuild **`grit-cli`** before re-running **`grit-bench`**; pass **`--grit target/dist/grit`** or rely on the harness preferring **`target/dist/grit`** over **`target/release/grit`**.
- **Trivial status RSS:** In an empty repo, peak RSS for **`grit status`** is measured with **`grit-bench measure-rss`** (fresh helper process, **`GRIT_BENCH_MEASURE_CMD='…/target/dist/grit status'`**) and/or **`VmHWM`** from **`/proc/<pid>/status`** over many cold starts (100+). Report median and max in bytes or KiB. **Target remains ≤ 8 MiB (8 388 608 bytes)**; the **`dist`** profile reaches **~9.4–9.8 MiB** on the factory VM (link tuning only). **Follow-up:** factory plan step under **[#2348](https://github.com/gitbutlerapp/grit/issues/2348)** — binary diet (feature-gate or lazy-load HTTP/jaq, reduce static init/relocations, allocator/startup mapping review) without slowing dashboard scenarios.
- **Fixtures** use synthetic repos at large file counts and deep history, built deterministically in a scratch directory.
- **Drivers** are CLI invocations of `grit` and `git` with equivalent semantics (not argv-for-argv compatibility). Dashboard scenarios `log-{size}` time a single **`grit log --limit 100`** against **`git log -100 --oneline`** on the synthetic history fixture. The production **campaign** matrix still has a separate **`log`** row that uses paged **`grit log --before=`** cursors on **git.git** (not the dashboard id).
- **Ratios** are **Grit median ÷ Git median** (stored in each scenario’s `ratio` field; hyperfine also records mean and spread). Values below `1.00×` mean Grit is faster; values above `2.00×` are highlighted as regressions worth investigating.
- **Spread** shows Grit’s standard deviation across timed runs (hyperfine `--min-runs` / warmup settings from the baseline capture).

**Revision-selection filters (`grit-bench revision-selection`).** Baseline **`grit-utils/baselines/revision-selection.json`** captures the deep-history network fixture with and without a reachable **commit-graph**: bounded **`--since`** counts, symmetric-range **left/right** counts, **`--author`** counting and **`-n 50`** listing (library **`lib-op`** drivers for rev-list-shaped workloads; **`grit log --author`** / **`grit revs --author`** for CLI rows), plus **`log-author-n50`** on the same fixture. Daily census registers the same workloads under **`REVISION_SELECTION_MATRIX_SCENARIOS`** (ids such as **`revs/count-since@deep-history`**, **`log/author-n50@deep-history-cg`**, and **`log/author@git-git`**). Pickaxe, **`--group-by author`**, **`--follow`**, and **`--cherry-pick`** scenarios join this baseline when their roadmap steps land.

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench revision-selection --format json --output grit-utils/baselines/revision-selection.json
```

### Maintenance (`grit-bench maintenance`)

Baseline **`grit-utils/baselines/maintenance-medium.json`** compares full **`grit gc`**, **`grit gc repack --bitmaps`**, and **`grit gc commit-graph`** against **`git gc`**, **`git repack -adb`**, and **`git commit-graph write --reachable --changed-paths`** on the deep-history network fixture with a worktree seeded with extra loose objects. Each hyperfine run uses **`prepare-maintenance`** to restore that worktree so timings are not skewed by mutating the same checkout across iterations.

Recorded on the factory VM when the baseline was captured (see `machine` / `tools` in the JSON). Roadmap item **15** targets **≤ 1.1×** Git on large-repo scenarios; maintenance work is tracked separately until repack and commit-graph close the gap:

| Scenario | Grit ÷ Git (committed baseline) | vs 1.1× bar | Follow-up |
| --- | ---: | --- | --- |
| `gc-medium` | **0.30×** | meets | — |
| `repack-medium` | **1.66×** | miss | [GitHub #1026](https://github.com/gitbutlerapp/grit/issues/1026) — pack build / bitmap sidecars |
| `commit-graph-write` | **1.17×** | miss | [#1026](https://github.com/gitbutlerapp/grit/issues/1026) — changed-path Bloom generation |

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench maintenance --format json --output /tmp/maintenance-medium.json
```

#### Full repack of a fresh clone

The maintenance suite also runs **`repack-cargo-shape`** (`grit repack` vs `git repack -ad`) and **`repack-force-cargo-shape`** (`grit repack --force` vs `git repack -adf`) on a synthetic cargo-shaped repository (`CargoShape::BENCH` in `grit-utils/src/repack_fixture.rs`: 8 000 commits over a nested crate layout with hot `Cargo.lock` / `CHANGELOG.md` files, imported with `git fast-import` and packed with `git repack -adf`, like a fresh clone).

Default repack copies stored entries, both whole objects and deltas with their base reference rewritten, and runs the delta search only on objects that are not already deltas against a packed base. `--force` searches every object in a streaming window. Both write the pack to a staging file. Measured with `/usr/bin/time` on [rust-lang/cargo](https://github.com/rust-lang/cargo) (23 565 commits, one 78 MB pack after `grit clone`, 4-vCPU Linux VM, git 2.43.0):

| Command | Before | After | `git` |
| --- | --- | --- | --- |
| `grit repack` vs `git repack -ad` | 2.5–4.0 s, 412 MB, 81.5 MB pack | 1.24–1.35 s, 194–197 MB, 81.4 MB pack | 1.32–1.41 s, 169 MB, 81.4 MB pack |
| `grit repack --force` vs `git repack -adf` | 393 s, 5.5 GB, 107 MB pack | 7.3–8.9 s, 459–480 MB, 45.7 MB pack | 6.9–7.0 s, 470–476 MB, 45.2 MB pack |
| `grit gc` vs `git gc` (already packed) | 0.33–0.37 s, 42 MB | 0.36–0.46 s, 42 MB | 1.53 s, 169 MB |

Before is the parent `main` commit; its forced row is that library's `force_recompress` path, which had no CLI flag yet.

The gate keeps this from regressing: `grit-utils/tests/repack_cargo_shape_budget.rs` builds `CargoShape::SMOKE` and requires both repack modes to stay within **2×** of git's peak memory (measured above each tool's `log -n 1` baseline, so debug-build code pages do not count) and, in optimized builds, within **2×** of git's time, with a pack no larger than **1.2×** git's. `grit-lib/tests/repack_delta_reuse_git_compat.rs` checks against the system `git` that default repack keeps every stored delta, `--force` stays within 1.2× of `git repack -adf`, and chains deeper than `pack.depth` are cut.

#### Repositories full of loose objects

Baseline **`grit-utils/baselines/gc-loose.json`** (`grit-bench gc-loose`) times full **`grit gc`** against **`git gc`**, with peak RSS, on repositories whose objects are all loose, restored from a pristine copy before every run (**`prepare-gc-loose`**). **`gc-loose-history`** is the deep-history network fixture with its pack exploded by `git unpack-objects`, so the delta search runs over every blob, tree and commit. **`gc-loose-wide`** is one commit of 200 000 small files (no delta work, many objects).

The same shape used to break the pack-objects path that serves fetches and clones and builds pushes and bundles ([#1883](https://github.com/gitbutlerapp/grit/issues/1883)). That path loaded every loose payload, ran a second 50-deep same-path pass and only deltified blobs. It now hands blobs, trees and commits to the same streaming window search as repack, with commits walked in date order and subtrees named by path. On rust-lang/cargo with its pack exploded into 194 636 loose objects (4 vCPU, git 2.43.0, `/usr/bin/time`):

| Command | Before | After | `git` |
| --- | --- | --- | --- |
| serve a full clone (`upload-pack`) | 195 s, 5.2 GB, 99.5 MB pack | 9.2 s, 207 MB, 39.9 MB pack | 17.0 s, 97 MB, 39.4 MB pack |
| `grit clone file://` vs `git clone file://` | 391 s, 5.2 GB, 102.5 MB pack | 11.5 s, 228 MB, 45.5 MB pack | 18.2 s, 95 MB, 39.3 MB pack |
| serve a full clone of the packed repository | 2.97 s, 360 MB | 2.63 s, 289 MB | 1.88 s, 166 MB |

`grit-lib/tests/pack_objects_loose_git_compat.rs` checks against the system `git` that a pack built from an all-loose history indexes under `git index-pack --strict`, deltifies most trees, and stays within **1.2×** of `git repack -adf`.

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench gc-loose --format json --output grit-utils/baselines/gc-loose.json
```

### Reproduce locally

Build release binaries, install [hyperfine](https://github.com/sharkdp/hyperfine), then run matching grit-bench scenarios from the repo root (see `grit-utils/README.md` for the current subcommands):

```bash
cargo build --profile dist -p grit-cli -p grit-utils

# Example: hot-path scenarios at L/H tiers (switch, pick, merge)
./target/release/grit-bench hot-paths --sizes 10000,100000 --format json --output /tmp/grit-bench-hot-paths.json
```

**LTO / dist profile (factory step 2370, factory VM):** Before profile tuning, a default **`release`** **`grit`** had **19 208 934** bytes of `.text` (GNU **`size`**, ~18.3 MiB), **28 642** dynamic relocations, **24 924 840** bytes on disk, **~11.7 MiB** median peak RSS on trivial **`grit status`** (**`measure-rss`**, empty repo), and **~184 s** clean **`cargo build --release -p grit-cli`**. The shipped **`dist`** profile cuts **`.text`** to **17 360 233** bytes (~16.5 MiB), relocations to **17 189**, on-disk size to **17 691 672** bytes, trivial status RSS to **~9.8 MiB** median (**`measure-rss`** **10 233 856** bytes; **`VmHWM`** **9996 KiB**), and clean **`cargo build --profile dist -p grit-cli`** to **~1.4×** the pre-LTO **`release`** build. **`opt-level = "s"`** was slower than **`opt-level = 3`** on every dashboard **10k** scenario on this VM, so **`dist`** uses **`3`**. RSS is still above the **8 MiB** goal; see the follow-up under **[#2348](https://github.com/gitbutlerapp/grit/issues/2348)**.

When roadmap item 2 adds more baseline JSON files on `main`, list their paths under `[benchmarks].baseline` in `content/docs/site.toml` and regenerate the site.

Compare a fresh run against the committed baseline (exit non-zero if any scenario ratio drifts beyond tolerance):

```bash
./target/release/grit-bench compare grit-utils/baselines/hot-paths-before.json /tmp/grit-bench-hot-paths.json --tolerance 0.10
```

### Hot-path acceptance (maintenance plan step 386)

Recorded on the factory VM after stacking pick/merge/stash perf work on `origin/main`. **Before** numbers are in `grit-utils/baselines/hot-paths-before.json` (2026-10-07); **after** in `grit-utils/baselines/hot-paths-after.json` (2026-10-08). Config isolation (`GIT_CONFIG_NOSYSTEM`, empty global config) applies to hot-path, suite `commit`, and related scenarios. FSMN variants: `grit-utils/baselines/hot-paths-after-fsmn.json`.

**Machine (2026-10-08 acceptance):** Intel Xeon (factory VM), **4** physical / **4** logical CPUs, **16 GiB** RAM, Linux **6.12.94+**, scratch filesystem **ext4**, `rustc` **1.99.0**, `git` **2.43.0**, release `grit` **0.5.1**, hyperfine **2.x** (see `machine` in each baseline JSON).

**Machine (2026-10-09 refresh):** same factory VM (Intel Xeon, **4** / **4** CPUs, **16 GiB** RAM, Linux **6.12.94+**, ext4 scratch), `rustc` **1.99.0**, `git` **2.43.0**, release `grit` **0.5.3** at **`01a254dc`**, hyperfine **2.0.0**. Committed baselines were regenerated with `./target/release/grit-bench {hot-paths,commit,all} --sizes 10000,100000` (default warmup / min-runs).

| Scenario | L before × | L 2026-10-08 × | L 2026-10-09 × | H before × | H 2026-10-08 × | H 2026-10-09 × | Notes |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | --- |
| `add` | — | **2.52** | **2.23** | — | **4.51** | **2.69** | `suite-after-LH.json` — improved at H |
| `commit` | — | **0.84** | **0.48** | — | **1.16** | **0.70** | `commit-after-LH.json` — faster vs Git at L/H |
| `switch` | 3.29 | 3.44 | **2.94** | 4.89 | 5.63 | **4.06** | H improved; L still &gt;2× |
| `switch-wide` | 3.70 | 2.29 | **3.42** | 26.5 | 2.98 | **22.54** | **H regression** — see refresh notes |
| `pick` | 1.41 | 1.14 | **1.48** | 5.50 | 4.04 | **5.22** | L slightly worse vs 2026-10-08 |
| `merge` | 2.61 | 2.18 | **2.86** | 3.97 | 3.83 | **3.74** | L ~32% worse ratio vs 2026-10-08 |
| `pick-series` | 12.9 | 11.2 | **11.94** | 21.7 | 17.7 | **18.56** | Still &gt;2×; 20× `grit pick` vs one `git cherry-pick` range |

**2026-10-09 refresh vs 2026-10-08 baselines:** `grit-bench compare` reports ratio drift beyond the default **0.10** absolute tolerance on most scenarios (many improved). Regressions **more than 20% worse** on the Grit/Git ratio: **`switch-wide-100000`** (2.98× → **22.54×**), **`switch-wide-10000`** (2.29× → 3.42×), **`merge-10000`**, **`pick-10000`**, **`pick-100000`**. **`status-dirty-100000`** remains ~**53×** (unchanged). Suite **status** and **add** ratios improved at L/H except dirty-100k.

| Scenario | Git ms | Grit ms | Old ratio | New ratio |
| --- | ---: | ---: | ---: | ---: |
| `status-dirty-10000` | 11.1 | 80.1 | 10.19× | 7.21× |
| `status-dirty-100000` | 113.0 | 5962.1 | 52.66× | 52.77× |
| `status-clean-10000` | 9.0 | 21.2 | 4.58× | 2.34× |
| `status-clean-100000` | 82.7 | 513.8 | 9.67× | 6.21× |
| `add-10000` | 21.4 | 47.8 | 2.52× | 2.23× |
| `add-100000` | 224.8 | 603.6 | 4.51× | 2.69× |
| `commit-10000` | 165.6 | 79.1 | 0.84× | 0.48× |
| `commit-100000` | 1665.7 | 1159.3 | 1.16× | 0.70× |
| `switch-10000` | 23.9 | 70.3 | 3.44× | 2.94× |
| `switch-100000` | 160.6 | 652.4 | 5.63× | 4.06× |
| `switch-wide-10000` | 45.3 | 155.3 | 2.29× | 3.42× |
| `switch-wide-100000` | 386.4 | 8711.8 | 2.98× | 22.54× |
| `pick-10000` | 115.9 | 171.5 | 1.14× | 1.48× |
| `pick-100000` | 194.0 | 1011.6 | 4.04× | 5.22× |
| `merge-10000` | 104.9 | 300.5 | 2.18× | 2.86× |
| `merge-100000` | 283.4 | 1060.6 | 3.83× | 3.74× |
| `pick-series-10000` | 129.4 | 1544.3 | 11.18× | 11.94× |
| `pick-series-100000` | 961.6 | 17843.4 | 17.65× | 18.56× |

**2026-10-11 status index lookups:** Replacing quadratic index scans during untracked/ignore classification dropped **`status-dirty-100000`** from ~**53×** to ~**2.5×** Git median time and **`status-dirty-10000`** from ~**7×** to ~**2×** (see refreshed `suite-after-LH.json`). A follow-up **size-mismatch** fast path on the status diff (Git `compare_stat_data`: different size ⇒ modified without re-reading blobs) brings **`status-dirty-10000`** to ~**1.94×** on the committed baseline. The shortcut applies only when the path's clean conversion is the identity. Under a line-ending rule, filter, `working-tree-encoding` or `ident`, a file of another size can clean to the indexed blob, so it is hashed, and a match adopts the new stat ([#2164](https://github.com/gitbutlerapp/grit/issues/2164)).

**Suite at L/H** (`grit-utils/baselines/suite-after-LH.json`): **`status-dirty-10000`** is below the **2×** bar; dirty status at 100k remains ~**2.2×** (within **3×**); clean status at L/H is still above **2×**.

**H/L acceptance (1.25× bar):** after-run medians fail the scaling check for **`switch`**, **`switch-wide`**, **`pick`**, **`merge`**, and **`pick-series`** (not only switch/pick/merge). **`commit`** and **`add`** at H are within 2× of Git but **`add`** at L/H both exceed 2× at 10k and 100k files.

Two back-to-back hot-path after runs (`hot-paths-after-run1.json` vs `run2.json`) agree within **10% relative** on every scenario except `merge-100000` and `pick-100000` (~**13–14%** relative spread). The default `grit-bench compare` tolerance is an **absolute** 0.10 on the ratio, which is tighter than ±10% relative for large ratios (e.g. pick-series).

### pick-range-20 (RSS gap)

**Scenario:** `grit pick base..topic` vs `git cherry-pick --no-edit base..topic` on the **pick-series** synthetic-**10000** fixture (20 commits on `topic`, one file touched per commit). Baseline: **`grit-utils/baselines/hot-paths-after.json`** (`pick-range-20`).

**Committed baseline (factory VM, release hyperfine, Git 2.43, 2026-10-24):** grit median **~94 ms** vs git **~118 ms** (**~0.79×** time); peak RSS **~1.42×** git (grit **~22.4 MiB** vs git **~15.8 MiB**). The earlier ~1.24× figure came from a `dist` (LTO, stripped) build with a smaller resident binary. Median time meets the **≤ 1.0×** goal; peak RSS remains above **≤ 1.2×** until shared replay/checkout work lands.

**Follow-up:** [GitHub issue #2977](https://github.com/gitbutlerapp/grit/issues/2977) — cross-cutting replay/merge work (index-hinted tree diffs in replay merge, scoped checkout for changed paths only, reduce peak RSS by avoiding full tree/index copies per step). Benefits **revert**, **merge**, and **rebase**, not pick-only code.

**Regression guard:** After merge-tree cache reuse and touched-path collection changes, **`pick-10000`** and **`merge-10000`** ratios on the same fixture did **not** regress vs the prior baseline (fresh hot-path run, 2026-10-23).

### Follow-up tracking (step 386)

**Binary diet for trivial status RSS (factory [#2348](https://github.com/gitbutlerapp/grit/issues/2348), after step 2370):** Reach **`grit status`** peak RSS **≤ 8 MiB** (8 388 608 bytes) in an empty repo without regressing any dashboard **10k** scenario. Candidates: feature-gate or lazy-load the HTTP (**`http-ureq`** / TLS) and **`jaq`** stacks, trim static initialization and dynamic relocations further, and review allocator/startup mappings. **Do not** lower the **8 MiB** target; step 2370 accepted **~9.4–9.8 MiB** as the link/profile outcome only.


No GitHub issues were filed from this factory run (`file_bug_report` is dogfooding-only). Track these against roadmap item 4:

- **L &gt;2×:** `switch`, `switch-wide`, `merge`, `pick-series`, `add` (10k and 100k), status at L/H.
- **H/L &gt;1.25×:** `switch`, `switch-wide`, `pick`, `merge`, `pick-series`.
- **Repeatability:** second hot-path after-run drift on `merge-100000` and `pick-100000` (~13–14% relative).
- **CLI gaps:** `grep`, `rebase`, `reset` (deferrals documented above).

**Criterion** (`cargo bench -p grit-lib --bench hot_paths`, release): see factory run log — `apply_stash_L` ~52 ms median; `pick_2000_paths_L` ~621 ms (library merge/checkout path, not CLI).

### Roadmap commands without a `grit` engine yet

| Git command | Grit CLI | How the plan covers it |
| --- | --- | --- |
| `grep` | none | Deferred — no porcelain driver in scope |
| `rebase` | none | `pick-series` grit-bench scenario (20× `grit pick`) as upper-bound proxy |
| `reset` | none | Deferred |
| `stash push` | `stash-push-pop` grit-bench scenario | `grit stash` + `grit stash pop` vs `git stash push` + `git stash pop` (see `stash-push-pop-LH.json`) |

### Refs

The **`grit-bench refs`** suite compares library drivers (`grit-bench refs-driver …`) to **`git for-each-ref`** and **`git update-ref --stdin`** on bare repositories with **100 000** branch refs. Fixtures are built once with system git and cached under the bench scratch directory: fully **packed** refs, **loose** refs only, and (when system git is **≥ 2.45**) a **reftable** layout. Scenarios list all refs, list the **`refs/heads/b05*`** glob (~1 000 refs), batch-**create** 100 000 refs into an empty-refs working copy, **update** 1 000 existing refs, and **delete** 1 000 refs; a prepare hook restores the fixture before each timed iteration. Initial packed/loose listing baselines are in **`grit-utils/baselines/refs-before.json`**; files-backend listing and batch **`update-ref`** optimizations on the 100k-ref fixtures are in **`grit-utils/baselines/refs-after.json`**; reftable timings (Git **≥ 2.45**) are in **`grit-utils/baselines/refs-after-reftable.json`**. The **`grit-utils/baselines/refs-after-994.json`** snapshot captures [`gc::update_refs`](https://docs.rs/grit-lib/latest/grit_lib/refs/gc/fn.update_refs.html) batching through one ref-store transaction ([`References::apply`](https://docs.rs/grit-lib/latest/grit_lib/refs/struct.References.html#method.apply) is the handle form) with a reused packed-refs cache instead of reparsing `packed-refs` per update.

On the **100k packed** fixture, **update 1 000 refs** (`update-ref-update-1k-refs-100k-packed`) dropped from **~184 s** Grit mean (4 485× git) to **~1.5 s** (**~36×** git at ~41 ms); the regression was repeated full `packed-refs` parses per update in the batch path, not transaction semantics.

The same fixtures serve two **`grit upload-pack`** scenarios against **`git upload-pack`**. The first is a protocol v2 **`ls-refs`** request with `peel`, `symrefs` and **`ref-prefix refs/heads/b05000`**, the request `git fetch origin b05000` sends (`upload-pack-ls-refs-prefix-*`). The second is the full v0 advertisement (`upload-pack-advertise-v0-*`). Both are read-only, so they run without the restore hook. Set **`GRIT_BENCH_REFS_ONLY_SERVE=1`** to run only these. Before [#1820](https://github.com/gitbutlerapp/grit/issues/1820), `ls-refs` listed and peeled every ref and only then applied the prefix: on the packed fixture it took **387 ms** for one ref (git **1.0 ms**), and **1.15 s** on the loose fixture (git **242 ms**). It now asks the ref store for the requested prefix only and peels only when the client asks. That brings it to **4.0 ms / 19.6 MB** packed and **42.5 ms / 12.9 MB** loose, where it beats git (**277 ms**) because the loose walk skips sibling ref files by name. The packed v0 advertisement went from **414 ms** to **117 ms** (git **16 ms**) because `packed-refs` peel lines replace tag reads. Before and after are in **`grit-utils/baselines/refs-serve-before.json`** and **`grit-utils/baselines/refs-serve-after.json`** (system `git 2.43.0`, one sitting).

```bash
cargo build --release -p grit-cli -p grit-utils
GRIT_BENCH_REFS_ONLY_SERVE=1 GRIT_BENCH_BRANCH=factory/ls-refs-prefix-first ./target/release/grit-bench --min-runs 10 refs --format json --output grit-utils/baselines/refs-serve-after.json
```

Criterion micro-benchmarks for the reftable backend: **`cargo bench -p grit-lib --bench refs_reftable`** (`for_each_ref_100k`, `prefix_seek`, `transaction_create_10k`).

```bash
cargo build --release -p grit-cli -p grit-utils
GRIT_BENCH_BRANCH=factory/refs-bench-774 ./target/release/grit-bench refs --format json --output grit-utils/baselines/refs-before.json
GRIT_BENCH_BRANCH=factory/files-refs-perf-782 ./target/release/grit-bench refs --format json --output grit-utils/baselines/refs-after.json
GRIT_BENCH_BRANCH=factory/reftable-refs-perf-783 GRIT_BENCH_REFS_ONLY=reftable ./target/release/grit-bench refs --git "$(command -v git)" --format json --output grit-utils/baselines/refs-after-reftable.json
```

Use a **≥ 2.45** system **`git`** (or an explicit `--git` path) when refreshing the reftable baseline; older git skips the reftable fixture.

Set `GRIT_BENCH_BRANCH` to the GitButler branch you are benchmarking so `grit_commit` in the JSON matches that branch tip (required on `gitbutler/workspace` checkouts). When `--grit` points at a binary built from a different commit than `grit-bench` itself (for example a before/after pair timed with one harness), set `GRIT_BENCH_REPORT_GRIT_COMMIT` to that binary's commit; every suite, including `grit-bench daily`, records it as `grit_commit`.

Driver **`for-each-ref`** output is checked against git byte-for-byte in `grit-utils` integration tests (1 000-ref fixture).

### Object reads

The **`grit-bench odb`** suite (library drivers in `grit-bench drive …` vs system **`git`**) measures cat-file batch reads, `rev-list --objects --all`, and `log -p` on a cached bare **git.git** clone and a repacked **100k-file / 1000-commit** synthetic repo. JSON reports include **peak RSS** per scenario (each workload measured in a fresh `grit-bench measure-rss` helper process via `getrusage(RUSAGE_CHILDREN)` on that run only) alongside wall time.

Reproduce:

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench odb --format json --output grit-utils/baselines/odb-read.json
```

**Measured outcomes:** the Object reads table below is generated from committed `grit-utils/baselines/odb-read.json` (hyperfine **≥5** runs per scenario, grit vs system git medians and peak RSS). Refresh with two invocations and `grit-bench compare run1.json run2.json --tolerance 0.10` before updating the baseline file and running `make docs`.

## grit-bench: deep-leaf worktree (issue #1017)

**`grit-bench deep-leaf`** builds the dogfood layout used on Windows release checks: many deep leaf directories with a few CRLF **`*.txt`** files each (via **`.gitattributes`**), a large **`large.bin`**, and tracked **`*.rs`** sources. Scenarios time clean **status**, dirty **status** after modifying **100** Rust files, and **diff** against system **`git`**. Use **`--smoke`** for a few hundred leaf dirs in CI; omit it for the full **~50k**-path production fixture.

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench deep-leaf --smoke --format json --output /tmp/deep-leaf-smoke.json
```

## grit-bench: dense-leaf worktree (issue #1041)

**`grit-bench dense-leaf`** builds the Windows dogfood layout with **many small files per leaf directory** (~**50** **`*.txt`** files under each of **~1000** deep prefixes, plus a few extra-deep chains). Scenarios time clean **status**, dirty **status** after modifying **100** text files, **diff**, and **add** against system **`git`**. The **add** scenario uses a hyperfine **`--prepare`** hook that runs **`git reset -q HEAD`** and re-dirties the same **100** files before each timed run. Use **`--smoke`** (~**4k** tracked paths) in CI; omit it for the full **~50k**-path production fixture. Release checks on Windows should compare medians to **`git`** and expect ratios within **3×** on warm-cache runs.

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench dense-leaf --smoke --format json --output /tmp/dense-leaf-smoke.json
```

## grit-bench: parallelism (threads=1 vs auto)

The **`grit-bench parallel`** suite measures how wall time scales when worker threads are pinned to **1** versus **auto** (unset `index.threads`, `checkout.workers`, and `pack.threads`, so system **git** and **grit** use all logical CPUs). Each scenario runs twice in the JSON baseline: ids end with **`-t1`** or **`-tauto`**. Hermetic git config matches other suites (`GIT_CONFIG_NOSYSTEM`, empty global config, `gc.auto=0`), with extra keys injected through **`GIT_CONFIG_COUNT`** for the one-thread runs. When the **`grit`** binary exposes a global **`--threads`** flag, the suite passes **`--threads=1`** or **`--threads=0`** (auto) in addition to those config keys.

Fixtures reuse the synthetic **100 000**-file repo (same shape as **`synthetic-100000`** / hot-path **H** tier). Scenarios:

| Workload | Git command / config | Grit |
| --- | --- | --- |
| Status dirty / clean | `git status -s` | `grit status` |
| Fresh checkout | `git checkout -f HEAD` after `git clone --no-checkout` | `grit-bench drive checkout-branch` (library checkout) |
| `file://` clone | `git clone` | `grit clone` |
| Wide switch | `git switch` out and back (~10% path delta) | `grit switch` |
| Index-pack | `git index-pack --threads=…` on the repacked `.pack` | `grit-bench drive index-pack` |
| Pack writing | `git pack-objects --all --stdout` | `grit-bench drive pack-build-all` |
| Rename diff | `git diff -M` (~2000 renames) | `grit-bench drive diff-rename-m` |

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench parallel --format json --output grit-utils/baselines/parallel-before.json
./target/release/grit-bench parallel --format json --output grit-utils/baselines/parallel-after.json
./target/release/grit-bench compare grit-utils/baselines/parallel-before.json grit-utils/baselines/parallel-after.json --tolerance 0.10
make docs
```

The committed **`parallel-before.json`** baseline captures **before** parallelism work (maintenance plan step 997). **`parallel-after.json`** (2026-10-12, roadmap item 15 close-out) captures the same scenarios after the full parallelism stack on **`origin/main`**, with **`git` 2.55.0**, release **`grit` 0.5.6**, hyperfine **1.19.0**, and **4** logical CPUs (see `machine` / `tools` in each JSON).

The suite reuses the repacked **100k** synthetic repo from **`GRIT_BENCH_ODB_CACHE`**, caches status and rename fixtures, and uses at least **7** hyperfine warmups and **10** timed runs per command (still overridable with global **`--warmup`** / **`--min-runs`**). Noisy scenarios warm **`git diff`** and **`pack-objects`** in **`--prepare`**. Fresh checkout warms on a **disposable** sibling clone, then re-clones **`--no-checkout`** into the timed destination (asserted empty) before each iteration. Generated site tables include **threads** in the scenario id suffix; compare **`-t1`** and **`-tauto`** rows for the same workload to see scaling. **`machine.logical_cores`** in the JSON records core count at capture time.

### Parallelism acceptance (100k synthetic repo, `-tauto`)

Roadmap item **15** bars **Grit median ÷ Git median ≤ 1.1×** on the large-repo scenarios below (auto threads for both tools). Ratios come from **`parallel-after.json`**.

| Scenario | Grit ÷ Git | vs 1.1× bar | Profile / notes |
| --- | ---: | --- | --- |
| **`parallel-status-dirty-100000-tauto`** | **2.67×** | miss | Down from ~**50×** in **`parallel-before.json`** after parallel stat/hash and untracked walks; remaining gap is mostly index refresh + untracked enumeration vs Git’s preload path on this fixture. |
| **`parallel-status-clean-100000-tauto`** | **3.28×** | miss | Clean worktrees still pay full index/untracked setup; Git **2.55** median ~**65 ms** vs Grit ~**213 ms**. |
| **`parallel-checkout-fresh-100000-tauto`** | **2.05×** | miss | Library checkout after **`clone --no-checkout`**; parallel file writers help but tree/blob materialization and index writes dominate. |
| **`parallel-clone-file-100000-tauto`** | **1.53×** | miss | **`file://`** clone including checkout; within **1.6×** but above the **1.1×** close-out bar. |
| **`parallel-index-pack-100000-tauto`** | **2.03×** | miss | In-memory **`drive index-pack`** on the repacked **100k** `.pack`; **~2.1×** Grit speedup **t1→tauto** but Git’s **`index-pack --threads=auto`** baseline is still faster. |

**Within bar elsewhere:** refreshed **`network.json`** **`clone-file-prod`** ≈ **0.95×** (deep-history **`file://`** clone). **`server-clone-http-compare-prod`** remains ≫ **1.2×** (upload-pack without reachability bitmaps — tracked under the bitmap/MIDX item, not parallelism).

### Thread scaling (`parallel-after.json`, Grit medians)

On **4** logical CPUs, pinning **`--threads=1`** vs **`--threads=0`** (auto) shows where pool policy pays off:

- **Status (dirty/clean, 100k):** ~**1.17–1.19×** faster at auto vs single-threaded Grit — parallel stat/hash helps, but the workload is still Git-limited on ratio.
- **`parallel-index-pack-100000`:** **~2.12×** faster **t1→tauto** (Grit **~713 ms** → **~337 ms**); largest win in the suite.
- **`parallel-switch-wide-100000`:** **~1.02×** **t1→tauto** — still **~13×** Git at auto (checkout/merge bound; not part of the 1.1× bar).
- **`parallel-pack-write-100000`:** **~3×** **t1→tauto** on Grit medians (delta encode + zlib pool); ratio vs Git at auto ~**75×** (Git streams a highly optimized **`pack-objects`**).

Refreshed companion baselines on the same VM capture **`status-*` / `add-*`** in **`suite-after-LH.json`**, **`switch-*`** in **`hot-paths-after.json`**, and **`clone-file-prod`** / **`server-clone-http-compare-prod`** in **`network.json`** (Git **2.55** for suite/network portions of the capture run).

Integration smoke (tiny fixtures, sixteen scenarios; requires **hyperfine** on `PATH`): `cargo test -p grit-utils parallel_scenario_smoke_end_to_end`.

## grit-bench: network (clone, fetch, push, ls-remote)

The **`grit-bench network`** suite compares **`grit`** and system **`git`** on the same remote URLs. Cached fixtures live under **`GRIT_BENCH_NETWORK_CACHE`** (default `/tmp/grit-bench-network-cache`):

| Fixture | Shape |
| --- | --- |
| **deep-history** | **≥50 000** commits and **≥20 000** tracked files (built once via `git fast-import`) |
| **many-refs** | deep-history plus **≥10 000** `refs/heads/bench-ref-*` branches |
| **large-blobs** | a small tree with multi‑MiB blobs for pack-heavy transfer |

Scenarios include **clone** over **`file://`**, **git http-backend** (local CGI), and **grit-http-server** smart HTTP; **incremental** and **no-op fetch** over all three transports; **push** over **`file://`** and **grit-http-server** (git http-backend push is omitted until the bench CGI harness supports receive-pack); **`git ls-remote`** vs **`grit remote refs`** on the many-refs fixture over all three transports; and a **server-side** comparison that times system **`git clone`** against **git http-backend** vs **grit-http-server** on the same published bare repo. Hermetic git config matches other suites (`GIT_CONFIG_NOSYSTEM`, empty global config).

```bash
cargo build --release -p grit-cli -p grit-http-server -p grit-utils
./target/release/grit-bench network --format json --output grit-utils/baselines/network.json
./target/release/grit-bench compare grit-utils/baselines/network.json /tmp/network-rerun.json --tolerance 0.10
make docs
```

Integration smoke (tiny fixtures, fifteen scenarios; requires **hyperfine** on `PATH`): `cargo test -p grit-utils network_scenario_smoke_end_to_end`.

Acceptance bars for the network performance pass are **≤1.2×** git wall time on client-side scenarios (clone, fetch, push, ls-remote over `file://` and smart HTTP). Compare the generated table ratios to that bar.

**Server-side clone (`server-clone-http-compare-*`):** system **`git clone`** against **grit-http-server** is expected to stay above **1.2×** until the upload-pack path uses **reachability bitmaps** during negotiation (same dependency as Git’s `upload-pack` on large, pack-served repos). Client-side grit and git both pay similar pack transfer cost; the gap is dominated by grit’s in-process upload-pack enumeration without bitmaps, not HTTP keep-alive or side-band framing. Track under the MIDX/bitmap roadmap item rather than masking with compatibility shortcuts.

Regenerate release binaries with **`cargo build --release -p grit-lib`** before **`grit-bench network`** so the timed **`grit`** binary includes library changes.

## grit-bench: history rewrite (`rewrite-history`)

The **`grit-bench rewrite-history`** suite times **`grit history`** reword, drop, fixup, and squash against system **`git history`** when the installed git is **≥2.54**, and always against a scripted **`git rebase`** / **`git rebase --onto`** / **`git reset --soft`** baseline on the same synthetic fixture. The fixture is generated once under **`GRIT_BENCH_SCRATCH`** (default `/tmp/grit-bench-scratch`): **~10 000** tracked files, **5 000** linear commits on `main`, and **50** local branches at varied depths (`bench/branch-*`). Each hyperfine iteration resets the worktree from a bare snapshot so refs stay isolated between runs.

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench rewrite-history --format json --output grit-utils/baselines/history.json
./target/release/grit-bench rewrite-history --smoke   # CI-sized fixture, one run per scenario
```

**Acceptance bars (factory VM, committed `history.json`):** deep **reword** ≤**1.0×** `git history reword` and ≥**5×** faster than the **`git rebase -i`** equivalent; other scenarios ≤**1.1×** `git history` when available. The baseline records the system **`git`** version string in `tools.git`.

**2026-10-17 status:** On **git 2.43.0** (no `git history` subcommand), git-side timings use the rebase/reset scripts above. **Reword**, **drop**, and **squash** meet the bars (reword ≈**0.03×** vs rebase, well above the **5×** grit-faster-than-rebase bar). **`history-fixup-500`** remains ≈**3.4×** git (`git commit --fixup` + **`rebase --autosquash`**) because grit performs a full three-way replay for every descendant commit while Git’s autosquash path stays on the C merge fast paths on this fixture. **Follow-up:** factory performance work on **`grit-lib` history fixup** replay (merge reuse / commit-graph scoped walks) — tracked as a separate factory item; do not weaken merge semantics to green the ratio.

## Real-world campaign

The **campaign** scorecard runs a shared CLI matrix on real-shaped fixtures: upstream **git.git** (pinned tag, repacked with bitmaps), a **chromium-class synthetic** repo (`400 000` files, `40 000` dirs, `100 000` commits, `50 000` refs, single pack + commit-graph), and optionally **linux** (`--with-linux` or `GRIT_BENCH_LINUX=1`). Scenario ids look like `git-dot-git/status-dirty` and `chromium-synth/switch`; reachability **rev-list** / **serve-clone** rows are owned by **ROADMAP item 14** (reachability bitmaps) and are not scored in this pass.

**Methodology.** Each scenario uses **hyperfine** (3 warmup, ≥5 timed runs) on the system **`git`** binary and a release **`grit`** build with hermetic git config. Ratios are **grit_median_ms / git_median_ms**; the bar is **≤1.1×** unless documented in [`grit-utils/perf-targets.toml`](https://github.com/gitbutlerapp/grit/blob/main/grit-utils/perf-targets.toml). Grit-side OOM or unsupported commands are recorded once (git timing still captured) so the suite can finish on the factory VM.

**Reproduce on the factory VM** (reuse `GRIT_BENCH_ODB_CACHE` after the first run):

```bash
cargo build --release -p grit-cli -p grit-utils
export GRIT_BENCH_ODB_CACHE=/tmp/grit-bench-odb-cache
./target/release/grit-bench campaign --no-network --format json --output grit-utils/baselines/campaign-after.json
./target/release/grit-bench campaign --ledger --from grit-utils/baselines/campaign-after.json
```

Committed baselines: **`campaign-before.json`** (2026-10-14, pre-campaign stack; chromium smoke-sized fixture) and **`campaign-after.json`** (2026-10-17, post-campaign stack on production synthetic). Both are listed under `[benchmarks].baseline` in `content/docs/site.toml`; the site merges duplicate scenario ids with **after** winning for campaign files.

**Before / after (representative sample — 15 of 30 overlapping scenario ids).** Values compare grit/git ratios from the two JSON reports (not absolute wall times). The full overlap set is in the committed JSON baselines and the generated campaign rows below. Ratios **below 1×** mean grit improved relative to git; ratios **above 1×** mean the gap widened (the after capture uses the production synthetic worktree while the before capture used a smaller smoke fixture for chromium-class scenarios).

| Scenario | Before ratio | After ratio | After / before |
| --- | ---: | ---: | ---: |
| `git-dot-git/clone-file` | 3.17× | 0.05× | 0.02× |
| `git-dot-git/diff-clean` | 7.76× | 2.46× | 0.32× |
| `git-dot-git/merge` | 14.66× | 6.55× | 0.45× |
| `git-dot-git/resolve` | 18.38× | 15.99× | 0.87× |
| `git-dot-git/status-clean` | 4.15× | 4.15× | 1.00× |
| `chromium-synth/status-dirty` | 1.87× | 2.10× | 1.13× |
| `git-dot-git/commit` | 1.84× | 2.42× | 1.32× |
| `git-dot-git/switch` | 4.37× | 5.89× | 1.35× |
| `git-dot-git/add` | 2.32× | 4.48× | 1.93× |
| `chromium-synth/add` | 2.07× | 2.88× | 1.39× |
| `chromium-synth/switch` | 5.16× | 13.43× | 2.60× |
| `chromium-synth/diff-clean` | 2.45× | 7.82× | 3.18× |
| `git-dot-git/cat` | 12.56× | 63.36× | 5.04× |
| `chromium-synth/merge` | 2.47× | 17.95× | 7.26× |
| `chromium-synth/commit` | 0.90× | 11.84× | 13.15× |

Full ledger (51 scenarios): `grit-bench campaign --ledger --from grit-utils/baselines/campaign-after.json` or see **`campaign-after.ledger.md`** in `grit-utils/baselines/`.

**Gap ledger.** Every scenario above **1.1×** (or grit-side skip) has a `documented` entry in `perf-targets.toml` pointing at **ROADMAP items 7, 11, 14, 19, and 21–30**. Item 14 rows remain `owned-by-item-14`.

**Superseded micro-baselines.** Per-command JSON files (`hot-paths-after.json`, `commit-after-LH.json`, etc.) remain for focused regressions; the **campaign** pair is the authoritative before/after scorecard for real-repository CLI work. Prefer updating campaign baselines when a change targets end-to-end porcelain on git.git or large synthetic worktrees.

## grit-bench: tree-sha256 tags

Scenario group **`tag-tree-sha256`** (`grit-bench tag-tree-sha256`) measures signed annotated tags with optional [tree-sha256](https://grit-scm.com/docs/library/tree-sha256.md/index.md) claims on **git-dot-git**, **chromium-synth**, and optionally **linux** worktrees (same fixture prep as the [real-world campaign](#real-world-campaign)).

| Scenario id | What it measures |
| --- | --- |
| `<fixture>/tag-verify-digest-no-signature` | `grit tag --verify --no-signature` vs a git-only `ls-tree` / `cat-file` / SHA-256 reproduction pipeline |
| `<fixture>/tag-verify-digest-reference` | Same grit verify path vs the reference `tree-sha256` binary when installed (skipped otherwise) |
| `<fixture>/tag-sign-tree-sha256` | `grit tag --sign --tree-sha256` total time vs `grit tag --sign` only (overhead should stay within recomputed hash time + **10 ms**) |
| `<fixture>/tag-verify-signed` | Full `grit tag --verify` vs `git verify-tag` on a signed tag carrying a tree-sha256 header |

**Acceptance bars (factory VM):** digest verify ≤ **1.2×** reference wall time when the reference tool is present, and ≤ the git-only pipeline on all three shapes; signed verify ≤ **max(signature, hash) + 10 ms** overhead on the sign scenario; peak RSS for each row is recorded in the committed JSON (`peak_rss_bytes`).

```bash
cargo build --release -p grit-cli -p grit-utils
export GRIT_BENCH_ODB_CACHE=/tmp/grit-bench-odb-cache
./target/release/grit-bench tag-tree-sha256 --format json --output grit-utils/baselines/tag-tree-sha256.json
# optional linux row:
GRIT_BENCH_LINUX=1 ./target/release/grit-bench tag-tree-sha256 --with-linux --format json --output grit-utils/baselines/tag-tree-sha256.json
```

Committed baseline: **`grit-utils/baselines/tag-tree-sha256.json`** (listed in `content/docs/site.toml`).

Criterion micro-benchmark: `cargo bench -p grit-lib --bench tree_sha256` (compare with `--baseline tree_sha256-before` when changing hot paths).

## grit-bench: repository fsck (`fsck-medium`)

Scenario **`fsck-medium`** (`grit-bench fsck`) runs **`grit fsck --no-dangling`** and **`git fsck --no-dangling`** on a cloned deep-history worktree (same fixture as other network-cache scenarios). Baseline: **`grit-utils/baselines/fsck-medium.json`**.

**Acceptance bar:** ≤**1.1×** git median wall time (same bar as other maintenance hot paths).

**2026-10-25 status:** Pack verification now streams each pack's delta trees depth-first on worker threads, checks every object as soon as it is inflated, and keeps only compact per-object facts; the connectivity walk re-reads trees instead of caching every entry. The committed baseline is **0.82×** git (~**156 ms** vs ~**191 ms** median, peak RSS ~**76 MiB** vs ~**42 MiB**), down from ~**9.4×** (~**1840 ms**). On full clones measured with `/usr/bin/time` on the factory VM:

| Repository | git fsck | grit fsck before | grit fsck now |
| --- | --- | --- | --- |
| rust-lang/cargo (195k objects, 80 MB pack) | 6.0 s, 232 MB | 3.5 s, 2.45 GB | **1.6 s, 277 MB** |
| git/git (422k objects, 322 MB pack) | 34.2 s, 513 MB | 74.3 s, 13.0 GB | **17.7 s, 638 MB** |

**Memory guard:** `grit-bench fsck` fails when grit's peak RSS exceeds both **2×** git's and git's plus **64 MiB** (the slack covers fixed per-thread overhead on small fixtures). The Rust test `grit-lib/tests/fsck_memory.rs` checks that fsck's peak does not grow with the uncompressed size of a delta-heavy pack.

```bash
cargo build --release -p grit-cli -p grit-utils
./target/release/grit-bench fsck --format json --output grit-utils/baselines/fsck-medium.json
```

## Checkout cancellation batches (2026-10-26)

Checkout now decides whether to use its worker pool before splitting planned jobs into cancellation batches. Batches grow from eight to 2,048 files, with a separate 1 MiB prepared-content threshold and bounded parallel read-ahead. Large files still stream; each write flush contains at most eight streams. Small loose-object cache reuse is limited to bulk checkout, while direct writes and a one-pass object walk avoid that bookkeeping.

On the factory VM (AMD EPYC 7763, four available logical CPUs, ext4, Rust 1.99.0, system Git 2.55.0), the six regression scenarios improve **13–25%** from main snapshot `12b52a55c` to `caf6f64fa`. The table reports median wall time and maximum child RSS, using release builds, one warmup, and 12 timed runs per tool for the shared checkout cases; additional daily A/B cases use ten runs per tool. All six orders are cycled so trial predecessors are balanced, including boundaries. Additional A/B rows show their after-session Git control; both Git controls are retained in the data. Small differences around 1 ms should be read against the identical-binary control spread.

| Scenario | Git ms | Grit before ms | Grit after ms | Peak RSS KiB before / after |
| --- | ---: | ---: | ---: | ---: |
| `switch/branch@synth-10k` | 46.9 | 33.4 | 33.6 | 22,824 / 22,736 |
| `restore/paths@synth-10k` | 21.2 | 21.5 | 21.7 | 22,572 / 22,652 |
| `reset/hard@synth-10k` | 42.6 | 45.6 | 45.0 | 25,196 / 25,208 |
| `worktree/add@synth-10k` | 325.3 | 426.0 | 320.7 | 26,164 / 27,124 |
| `merge/clean@synth-10k` | 169.0 | 184.1 | 145.6 | 28,148 / 27,400 |
| `pick/one@synth-10k` | 182.5 | 185.1 | 161.8 | 27,620 / 27,992 |
| `switch/branch@synth-100k` | 241.6 | 107.3 | 105.5 | 54,692 / 54,816 |
| `switch/delete-heavy@wide-48k` | 227.6 | 207.3 | 196.3 | 35,376 / 35,784 |
| `switch/large-blob@big-asset` | 843.2 | 478.6 | 474.2 | 18,648 / 18,508 |
| `merge/clean@synth-100k` | 458.2 | 274.7 | 234.8 | 59,492 / 60,240 |
| `revert/one@synth-10k` | 182.6 | 194.7 | 158.3 | 28,176 / 28,240 |
| `pick/one@synth-100k` | 285.5 | 347.1 | 296.4 | 58,140 / 58,496 |
| `clone/checkout@synth-100k` | 7077.6 | 5933.7 | 4473.5 | 79,072 / 81,128 |
| `merge/content-both-sides@wide-20k` | 2113.5 | 3020.0 | 2679.7 | 112,440 / 114,052 |
| `clone/local@synth-10k` | 1052.4 | 1364.6 | 1264.7 | 31,172 / 31,804 |
| `clone/http@synth-10k` | 1139.6 | 1415.9 | 1308.9 | 31,964 / 32,260 |
| `clone/bare@synth-10k` | 711.4 | 795.3 | 796.8 | 24,300 / 24,680 |
| `clone/shallow-depth1@synth-10k` | 412.5 | 504.4 | 398.2 | 27,804 / 28,176 |
| `clone/shallow-depth1-http@synth-10k` | 422.6 | 514.6 | 408.7 | 27,600 / 28,176 |
| `worktree/remove@synth-10k` | 204.1 | 174.1 | 174.6 | 13,620 / 13,496 |
| `worktree/list@synth-10k` | 2.9 | 3.7 | 3.6 | 12,648 / 12,752 |
| `reset/hard-delete-heavy@wide-48k` | 380.8 | 421.3 | 410.3 | 59,296 / 62,828 |
| `reset/mixed@synth-10k` | 15.7 | 18.9 | 18.7 | 20,208 / 20,480 |
| `clone/full@git-git` | 20627.0 | 10821.3 | 10643.9 | 185,648 / 184,988 |
| `clone/shallow@git-git` | 2147.0 | 2312.6 | 2241.0 | 73,308 / 72,148 |
| `clone/blobless@git-git` | 12120.1 | 9740.9 | 9656.4 | 93,232 / 88,988 |

Memory guards cover 256 distinct 512 KiB blobs under serial and parallel preparation and an already-cached 64 MiB blob. Prepared-buffer growth stays below 16 MiB; checking cached size before copying reduces the cached-large-blob guard's additional RSS from 65,176 KiB in the demonstrated failing version to 172 KiB. Round-trips use system Git (`write-tree`, clean status, and strict fsck). Cancellation and configured pool thresholds have Rust guards.

Allocation profiling of `pick/one@synth-10k` drops from 312,599 to 287,497 calls (8.0%); reported peak heap is 7.34M in both. Instrumented wall time and RSS are excluded from the table.

Data: **`grit-utils/baselines/daily-checkout-chunks-before.json`**, **`grit-utils/baselines/daily-checkout-chunks-after.json`**. The detailed investigation, exact commands, samples, historical endpoint/direct-parent probes, GNU time output, syscall summaries, controls and validation limitations are in **`grit-utils/baselines/checkout-chunks-4567-audit.json`** and **`grit-utils/baselines/checkout-chunks-4567.md`**. The first bad commit was `63ea7f3ef`: eight-file cancellation flushes repeatedly failed the parallelism threshold. Its cancellation behavior remains supported.

### Results

| | |
| --- | --- |
| Git | git version 2.43.0 |
| Grit | grit 0.5.15 |
| CPU | Intel(R) Xeon(R) Processor |
| OS | linux (Linux 6.12.94+) |
| Recorded | 2026-10-22 00:11:00.01375475 |

## Summary by operation

| Operation | Scenarios | Median Grit / Git | Worst Grit / Git |
| --- | ---: | ---: | ---: |
| add | 11 | 1.65× | **6.74×** |
| add-ignorecase-1682 | 16 | 1.35× | **2.56×** |
| bitmaps | 12 | **9.57×** | **408.18×** |
| blame | 8 | 0.85× | **16.88×** |
| branch_contains | 30 | 0.17× | **76.79×** |
| cat | 4 | 1.69× | **61.93×** |
| clap-globals-5562 | 24 | 1.51× | 1.71× |
| clone | 2 | 0.46× | 0.46× |
| commit | 4 | 1.73× | **11.15×** |
| conflicts | 9 | 0.68× | **5.41×** |
| delta_encode | 2 | — | — |
| diff | 6 | **4.03×** | **7.78×** |
| filtered-revision-count | 22 | 1.54× | **3.07×** |
| filtered-revision-count-range | 14 | 1.61× | **47.72×** |
| git-git-headlines | 4 | **2.15×** | **6.03×** |
| hash | 2 | 0.62× | 0.63× |
| history | 9 | 0.86× | **5.61×** |
| ignore | 21 | **2.77×** | **7.22×** |
| large-blob-materialization | 5 | 0.73× | 0.85× |
| log | 6 | 1.40× | **11.98×** |
| maintenance | 12 | 0.57× | **38.97×** |
| merge | 11 | 1.10× | **62.06×** |
| merge-base | 3 | 0.00× | **24.39×** |
| midx | 3 | — | — |
| network | 6 | **7.89×** | **47.55×** |
| network-clone | 3 | 1.02× | 1.85× |
| network-fetch | 30 | 1.26× | **27.57×** |
| network-ls-remote | 3 | 1.00× | 1.60× |
| network-push | 13 | **5.93×** | **36.93×** |
| network-server | 1 | **30.67×** | **30.67×** |
| object_reads | 8 | **234.49×** | **4569.02×** |
| odb | 3 | 1.39× | **3.43×** |
| odb_backend | 6 | **8.38×** | **98.20×** |
| odb_read_parallel | 4 | — | — |
| parallel | 16 | **2.82×** | **264.04×** |
| pick | 9 | **5.88×** | **1461.52×** |
| primary-store-7286 | 14 | 0.94× | 1.24× |
| ref-query | 3 | **14.33×** | **32.46×** |
| refs | 35 | 1.12× | **7.22×** |
| reset | 7 | 1.12× | 1.78× |
| resolve | 2 | **17.00×** | **17.34×** |
| resolve-cat | 2 | **2.31×** | **2.36×** |
| restore | 2 | 1.29× | 1.33× |
| revert | 2 | **777.59×** | **1527.28×** |
| revision-selection | 2 | **2.25×** | **2.49×** |
| revs | 10 | **2.26×** | **206.68×** |
| revs-startup-4992 | 24 | 1.63× | **2.10×** |
| startup-relocations | 80 | 1.83× | **3.59×** |
| stash | 4 | **2.65×** | **18.52×** |
| status | 19 | 1.20× | **5.99×** |
| switch | 8 | **3.48×** | **13.52×** |
| tag-tree-sha256 | 6 | 1.07× | **13.15×** |
| tree-metadata | 24 | 1.25× | 1.98× |
| tree-metadata-criterion | 4 | — | — |
| worktree | 5 | **2.77×** | **3.70×** |
| worktree-symlinks | 8 | 0.17× | 0.40× |

### add

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `add-big-file-1024m` | big-file-1024m | 19,264 | 7,355 | 0.38× | ±123 ms |
| `add-big-file-256m` | big-file-256m | 2,271 | 933 | 0.41× | ±1.98 ms |
| `chromium-synth/add` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 1,043 | 3,012 | **2.89×** | ±78.4 ms |
| `git-dot-git/add` | git-dot-git | 11.8 | 46.8 | **3.98×** | ±12.3 ms |
| `large-blob-memory-add-1024m` | incompressible-1024m | 44,900 | 840 | 0.02× | ±0.00 ms |
| `large-blob-memory-add-128m` | incompressible-128m | 390 | 90.0 | 0.23× | ±0.00 ms |
| `add-dot-many-dirs-4000` | many-dirs-4000 | 33.8 | 60.9 | 1.80× | ±1.56 ms |
| `add-initial-monorepo-20000` | monorepo-untracked-20000 | 1,027 | 550 | 0.54× | ±21.9 ms |
| `add-dot-nested-dirs-49000` | nested-dirs-49000 | 246 | 553 | **2.25×** | ±2.12 ms |
| `add-10000` | synthetic-10000 | 20.9 | 141 | **6.74×** | ±8.86 ms |
| `add-100000` | synthetic-100000 | 207 | 343 | 1.65× | ±17.5 ms |

### add-ignorecase-1682

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `add-ignorecase-1682-macos-path-config-add-1-path` | 48,020 tracked files, 98 modified (core.ignorecase + core.precomposeunicode) | 28.0 | 53.5 | 1.91× | ±0.00 ms |
| `add-ignorecase-1682-macos-path-config-add-10-paths` | 48,020 tracked files, 98 modified (core.ignorecase + core.precomposeunicode) | 31.2 | 50.6 | 1.62× | ±0.00 ms |
| `add-ignorecase-1682-macos-path-config-add-98-paths` | 48,020 tracked files, 98 modified (core.ignorecase + core.precomposeunicode) | 72.6 | 60.8 | 0.84× | ±0.00 ms |
| `add-ignorecase-1682-macos-path-config-add-add-a` | 48,020 tracked files, 98 modified (core.ignorecase + core.precomposeunicode) | 56.4 | 108 | 1.91× | ±0.00 ms |
| `add-ignorecase-1682-macos-path-config-add-dry-run` | 48,020 tracked files, 98 modified (core.ignorecase + core.precomposeunicode) | 59.6 | 87.6 | 1.47× | ±0.00 ms |
| `add-ignorecase-1682-macos-path-config-add-u` | 48,020 tracked files, 98 modified (core.ignorecase + core.precomposeunicode) | 39.2 | 66.8 | 1.70× | ±0.00 ms |
| `add-ignorecase-1682-macos-path-config-commit-m-x-after-add-a` | 48,020 tracked files, 98 modified (core.ignorecase + core.precomposeunicode) | 54.2 | 139 | **2.56×** | ±0.00 ms |
| `add-ignorecase-1682-macos-path-config-status-control` | 48,020 tracked files, 98 modified (core.ignorecase + core.precomposeunicode) | 39.4 | 46.4 | 1.18× | ±0.00 ms |
| `add-ignorecase-1682-default-add-1-path` | 48,020 tracked files, 98 modified (default config) | 25.5 | 30.5 | 1.20× | ±0.00 ms |
| `add-ignorecase-1682-default-add-10-paths` | 48,020 tracked files, 98 modified (default config) | 28.0 | 28.7 | 1.02× | ±0.00 ms |
| `add-ignorecase-1682-default-add-98-paths` | 48,020 tracked files, 98 modified (default config) | 70.4 | 36.5 | 0.52× | ±0.00 ms |
| `add-ignorecase-1682-default-add-add-a` | 48,020 tracked files, 98 modified (default config) | 50.8 | 83.0 | 1.63× | ±0.00 ms |
| `add-ignorecase-1682-default-add-dry-run` | 48,020 tracked files, 98 modified (default config) | 53.9 | 59.1 | 1.10× | ±0.00 ms |
| `add-ignorecase-1682-default-add-u` | 48,020 tracked files, 98 modified (default config) | 34.0 | 41.5 | 1.22× | ±0.00 ms |
| `add-ignorecase-1682-default-commit-m-x-after-add-a` | 48,020 tracked files, 98 modified (default config) | 49.8 | 111 | **2.22×** | ±0.00 ms |
| `add-ignorecase-1682-default-status-control` | 48,020 tracked files, 98 modified (default config) | 38.3 | 43.2 | 1.13× | ±0.00 ms |

### bitmaps

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `git-dot-git/rev-list-count` | git-dot-git | 54.9 | 6,253 | **113.95×** | ±72.0 ms |
| `git-dot-git/rev-list-count-objects` | git-dot-git | 14.9 | 6,073 | **408.18×** | ±43.4 ms |
| `git-dot-git/serve-clone` | git-dot-git | 289 | 0.00 | 0.00× | ±0.00 ms |
| `rev-list-count-git.git` | git.git | 57.4 | 686 | **11.95×** | ±3.36 ms |
| `rev-list-count-git.git-pack-reuse` | git.git | 66.9 | 1,198 | **17.92×** | ±19.5 ms |
| `rev-list-count-objects-git.git` | git.git | 16.9 | 714 | **42.12×** | ±11.9 ms |
| `rev-list-count-objects-git.git-pack-reuse` | git.git | 17.7 | 1,082 | **61.21×** | ±95.2 ms |
| `serve-clone-blobless-git.git` | git.git | 304 | 2,186 | **7.19×** | ±68.4 ms |
| `serve-clone-blobless-no-bitmaps-git.git` | git.git | 2,844 | 8,823 | **3.10×** | ±156 ms |
| `serve-clone-git.git` | git.git | 660 | 4,246 | **6.43×** | ±123 ms |
| `serve-clone-git.git-pack-reuse` | git.git | 684 | 0.00 | 0.00× | ±0.00 ms |
| `serve-clone-no-bitmaps-git.git` | git.git | 4,019 | 11,974 | **2.98×** | ±253 ms |

### blame

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `blame-large-tail-full` | blame-large-5000x150 | 73.7 | 37.2 | 0.51× | ±3.11 ms |
| `blame-large-tail-range` | blame-large-5000x150 | 8.10 | 7.12 | 0.88× | ±0.37 ms |
| `blame-deep-200` | blame-linear-200 | 8.23 | 4.67 | 0.57× | ±0.30 ms |
| `blame-deep-2000` | blame-linear-2000 | 67.4 | 58.7 | 0.87× | ±2.93 ms |
| `blame-merge-heavy-full` | blame-merges-3000x150 | 104 | 85.6 | 0.83× | ±9.35 ms |
| `blame-large-mid-churn-full` | blame-mid-3500x30 | 8.88 | 12.6 | 1.42× | ±1.39 ms |
| `chromium-synth/blame` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 232 | 124 | 0.53× | ±2.92 ms |
| `git-dot-git/blame` | git-dot-git | 194 | 3,279 | **16.88×** | ±45.3 ms |

### branch_contains

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `branch/contains-1-before@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34.4 | 2,642 | **76.79×** | ±15.7 ms |
| `branch/contains-1-reviewed@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34.4 | 14.7 | 0.43× | ±0.17 ms |
| `branch/contains-1@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34.4 | 14.7 | 0.43× | ±0.39 ms |
| `branch/contains-1000-before@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,609 | 2,637 | 0.08× | ±0.00 ms |
| `branch/contains-1000-reviewed@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,609 | 228 | 0.01× | ±0.00 ms |
| `branch/contains-1000@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,609 | 230 | 0.01× | ±0.00 ms |
| `branch/contains-before@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 199 | 2,612 | **13.10×** | ±5.03 ms |
| `branch/contains-json-1-before@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 35.5 | 2,610 | **73.48×** | ±5.16 ms |
| `branch/contains-json-1-reviewed@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 35.5 | 14.7 | 0.41× | ±0.04 ms |
| `branch/contains-json-1@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 35.5 | 14.4 | 0.40× | ±0.24 ms |
| `branch/contains-json-1000-before@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,185 | 2,630 | 0.08× | ±0.00 ms |
| `branch/contains-json-1000-reviewed@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,185 | 250 | 0.01× | ±0.00 ms |
| `branch/contains-json-1000@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,185 | 233 | 0.01× | ±0.00 ms |
| `branch/contains-json-before@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 199 | 2,615 | **13.12×** | ±8.44 ms |
| `branch/contains-json-reviewed@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 199 | 34.3 | 0.17× | ±0.09 ms |
| `branch/contains-json-unlimited-before@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,140 | 2,622 | 0.08× | ±0.00 ms |
| `branch/contains-json-unlimited-reviewed@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,140 | 2,367 | 0.07× | ±0.00 ms |
| `branch/contains-json-unlimited@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,140 | 2,389 | 0.07× | ±0.00 ms |
| `branch/contains-json@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 199 | 33.9 | 0.17× | ±0.26 ms |
| `branch/contains-reviewed@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 199 | 33.7 | 0.17× | ±0.46 ms |
| `branch/contains-unlimited-before@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,405 | 2,626 | 0.08× | ±0.00 ms |
| `branch/contains-unlimited-reviewed@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,405 | 2,394 | 0.07× | ±0.00 ms |
| `branch/contains-unlimited@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 34,405 | 2,352 | 0.07× | ±0.00 ms |
| `branch/contains@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 199 | 33.7 | 0.17× | ±0.27 ms |
| `branch/unfiltered-json-unfiltered-before@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 159 | 183 | 1.15× | ±0.01 ms |
| `branch/unfiltered-json-unfiltered-reviewed@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 159 | 180 | 1.13× | ±1.03 ms |
| `branch/unfiltered-json-unfiltered@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 159 | 181 | 1.14× | ±0.13 ms |
| `branch/unfiltered-unfiltered-before@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 162 | 135 | 0.84× | ±0.41 ms |
| `branch/unfiltered-unfiltered-reviewed@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 162 | 135 | 0.83× | ±0.32 ms |
| `branch/unfiltered-unfiltered@git-git-10k-refs` | git-git-2035-10k-unique-packed-tips | 162 | 134 | 0.83× | ±0.43 ms |

### cat

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/cat` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 13.8 | 42.7 | **3.09×** | ±4.86 ms |
| `git-dot-git/cat` | git-dot-git | 1.68 | 104 | **61.93×** | ±3.29 ms |
| `large-blob-memory-cat-1024m-loose` | incompressible-1024m-loose | 910 | 210 | 0.23× | ±20.0 ms |
| `large-blob-memory-cat-1024m-packed` | incompressible-1024m-packed | 860 | 247 | 0.29× | ±28.9 ms |

### clap-globals-5562

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `clap-globals-5562-log-author-n50-deep-history-after` | 50,000 commits, 20,000 files; no commit-graph | 1.50 | 2.56 | 1.70× | ±0.18 ms |
| `clap-globals-5562-log-author-n50-deep-history-before` | 50,000 commits, 20,000 files; no commit-graph | 1.50 | 2.57 | 1.71× | ±0.16 ms |
| `clap-globals-5562-branch-create-many-refs-after` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 1.40 | 2.23 | 1.59× | ±0.26 ms |
| `clap-globals-5562-branch-create-many-refs-before` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 1.40 | 2.25 | 1.61× | ±0.29 ms |
| `clap-globals-5562-branch-delete-many-refs-after` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 4.16 | 4.22 | 1.02× | ±0.55 ms |
| `clap-globals-5562-branch-delete-many-refs-before` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 4.16 | 4.40 | 1.06× | ±0.58 ms |
| `clap-globals-5562-branch-list-json-many-refs-after` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 1.32 | 1.90 | 1.44× | ±0.31 ms |
| `clap-globals-5562-branch-list-json-many-refs-before` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 1.32 | 1.91 | 1.44× | ±0.16 ms |
| `clap-globals-5562-branch-list-many-refs-after` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 1.14 | 1.73 | 1.52× | ±0.12 ms |
| `clap-globals-5562-branch-list-many-refs-before` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 1.14 | 1.75 | 1.53× | ±0.13 ms |
| `clap-globals-5562-branch-list-markdown-many-refs-after` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 1.36 | 1.90 | 1.40× | ±0.12 ms |
| `clap-globals-5562-branch-list-markdown-many-refs-before` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 1.36 | 1.93 | 1.42× | ±0.36 ms |
| `clap-globals-5562-branch-merged-many-refs-after` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 1.26 | 1.92 | 1.52× | ±0.21 ms |
| `clap-globals-5562-branch-merged-many-refs-before` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 1.26 | 1.96 | 1.55× | ±0.42 ms |
| `clap-globals-5562-tag-list-many-refs-after` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 2.73 | 3.81 | 1.40× | ±0.18 ms |
| `clap-globals-5562-tag-list-many-refs-before` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 2.73 | 3.84 | 1.41× | ±0.29 ms |
| `clap-globals-5562-version-after` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 0.85 | 1.14 | 1.35× | ±0.11 ms |
| `clap-globals-5562-version-before` | clone of the 50,000-commit deep-history fixture with 10,000 packed remote-tracking refs (one local branch) | 0.85 | 1.11 | 1.31× | ±0.13 ms |
| `clap-globals-5562-config-get-small-after` | small synthetic worktree | 0.98 | 1.59 | 1.61× | ±0.11 ms |
| `clap-globals-5562-config-get-small-before` | small synthetic worktree | 0.98 | 1.61 | 1.63× | ±0.12 ms |
| `clap-globals-5562-status-clean-small-after` | small synthetic worktree | 1.48 | 2.27 | 1.54× | ±0.16 ms |
| `clap-globals-5562-status-clean-small-before` | small synthetic worktree | 1.48 | 2.29 | 1.55× | ±0.15 ms |
| `clap-globals-5562-branch-list-many-refs-loose-after` | the many-refs fixture with loose refs | 1.17 | 1.73 | 1.48× | ±0.13 ms |
| `clap-globals-5562-branch-list-many-refs-loose-before` | the many-refs fixture with loose refs | 1.17 | 1.76 | 1.51× | ±0.18 ms |

### clone

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `large-blob-memory-clone-1024m-loose` | incompressible-1024m-loose | 740 | 330 | 0.45× | ±17.3 ms |
| `large-blob-memory-clone-1024m-packed` | incompressible-1024m-packed | 710 | 330 | 0.46× | ±34.6 ms |

### commit

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/commit` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 1,669 | 18,605 | **11.15×** | ±5,183 ms |
| `git-dot-git/commit` | git-dot-git | 68.8 | 167 | **2.42×** | ±20.9 ms |
| `commit-10000` | synthetic-10000 | 191 | 197 | 1.03× | ±12.5 ms |
| `commit-100000` | synthetic-100000 | 1,738 | 1,241 | 0.71× | ±94.6 ms |

### conflicts

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `merge-conflict-continue-1000` | synthetic-1000 | 53.3 | 117 | **2.19×** | ±52.7 ms |
| `merge-no-ff-1000` | synthetic-1000 | 28.4 | 19.2 | 0.68× | ±1.03 ms |
| `pick-range-20-1000` | synthetic-1000 | 29.7 | 46.8 | 1.58× | ±5.50 ms |
| `revert-mainline-1000` | synthetic-1000 | 38.3 | 23.9 | 0.62× | ±4.02 ms |
| `revert-single-1000` | synthetic-1000 | 38.4 | 23.2 | 0.60× | ±1.71 ms |
| `merge-conflict-continue` | synthetic-10000 | 196 | 1,061 | **5.41×** | ±23.0 ms |
| `merge-no-ff` | synthetic-10000 | 110 | 82.5 | 0.75× | ±3.12 ms |
| `revert-mainline` | synthetic-10000 | 111 | 69.9 | 0.63× | ±4.27 ms |
| `revert-single` | synthetic-10000 | 113 | 66.4 | 0.59× | ±2.11 ms |

### delta_encode

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `delta-encode-binary-256k` | synthetic-256k-binary | — | 0.36 | — | ±0.00 ms |
| `delta-encode-text-256k` | synthetic-256k-text | — | 0.54 | — | ±0.01 ms |

### diff

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/diff-clean` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 130 | 1,009 | **7.78×** | ±10.1 ms |
| `chromium-synth/diff-dirty` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 140 | 1,031 | **7.37×** | ±31.3 ms |
| `git-dot-git/diff-clean` | git-dot-git | 9.52 | 29.0 | **3.05×** | ±10.3 ms |
| `git-dot-git/diff-dirty` | git-dot-git | 13.6 | 68.2 | **5.00×** | ±14.5 ms |
| `diff-10000` | synthetic-10000 | 5.79 | 13.4 | **2.31×** | ±4.44 ms |
| `diff-100000` | synthetic-100000 | 39.9 | 106 | **2.64×** | ±6.74 ms |

### filtered-revision-count

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `filtered-log-author-n50-deep-history-after` | 50,000 commits, 20,000 files; no commit-graph | 2.11 | 3.57 | 1.69× | ±0.04 ms |
| `filtered-log-author-n50-deep-history-before` | 50,000 commits, 20,000 files; no commit-graph | 2.11 | 3.57 | 1.69× | ±0.05 ms |
| `filtered-revs-author-n50-deep-history-after` | 50,000 commits, 20,000 files; no commit-graph | 2.17 | 3.63 | 1.67× | ±0.06 ms |
| `filtered-revs-author-n50-deep-history-before` | 50,000 commits, 20,000 files; no commit-graph | 2.17 | 3.63 | 1.67× | ±0.06 ms |
| `filtered-revs-count-author-deep-history-after` | 50,000 commits, 20,000 files; no commit-graph | 92.7 | 120 | 1.30× | ±3.49 ms |
| `filtered-revs-count-author-deep-history-before` | 50,000 commits, 20,000 files; no commit-graph | 92.7 | 285 | **3.07×** | ±6.23 ms |
| `filtered-revs-count-since-deep-history-after` | 50,000 commits, 20,000 files; no commit-graph | 2.06 | 2.89 | 1.40× | ±0.09 ms |
| `filtered-revs-count-since-deep-history-before` | 50,000 commits, 20,000 files; no commit-graph | 2.06 | 2.87 | 1.39× | ±0.07 ms |
| `filtered-revs-left-right-count-deep-history-after` | 50,000 commits, 20,000 files; no commit-graph | 2.76 | 5.46 | 1.98× | ±0.06 ms |
| `filtered-revs-left-right-count-deep-history-before` | 50,000 commits, 20,000 files; no commit-graph | 2.76 | 5.43 | 1.97× | ±0.07 ms |
| `filtered-log-author-n50-deep-history-cg-after` | 50,000 commits, 20,000 files; system Git commit-graph | 2.38 | 3.66 | 1.54× | ±0.09 ms |
| `filtered-log-author-n50-deep-history-cg-before` | 50,000 commits, 20,000 files; system Git commit-graph | 2.38 | 3.66 | 1.54× | ±0.12 ms |
| `filtered-revs-author-n50-deep-history-cg-after` | 50,000 commits, 20,000 files; system Git commit-graph | 2.49 | 3.79 | 1.52× | ±0.14 ms |
| `filtered-revs-author-n50-deep-history-cg-before` | 50,000 commits, 20,000 files; system Git commit-graph | 2.49 | 3.77 | 1.51× | ±0.09 ms |
| `filtered-revs-count-author-deep-history-cg-after` | 50,000 commits, 20,000 files; system Git commit-graph | 90.6 | 124 | 1.36× | ±2.02 ms |
| `filtered-revs-count-author-deep-history-cg-before` | 50,000 commits, 20,000 files; system Git commit-graph | 90.6 | 172 | 1.90× | ±6.90 ms |
| `filtered-revs-count-author-git-git-after` | git/git v2.47.0 (default daily fixture); no commit-graph | 479 | 386 | 0.80× | ±5.06 ms |
| `filtered-revs-count-author-git-git-before` | git/git v2.47.0 (default daily fixture); no commit-graph | 479 | 918 | 1.91× | ±14.7 ms |
| `filtered-revs-count-git-git-after` | git/git v2.47.0 (default daily fixture); no commit-graph | 382 | 534 | 1.40× | ±9.87 ms |
| `filtered-revs-count-git-git-before` | git/git v2.47.0 (default daily fixture); no commit-graph | 382 | 538 | 1.41× | ±24.8 ms |
| `filtered-revs-count-grep-git-git-after` | git/git v2.47.0 (default daily fixture); no commit-graph | 486 | 366 | 0.75× | ±5.84 ms |
| `filtered-revs-count-grep-git-git-before` | git/git v2.47.0 (default daily fixture); no commit-graph | 486 | 887 | 1.82× | ±22.7 ms |

### filtered-revision-count-range

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `filtered-count-range-review-10` | 50,000 commits / 20,000 files; with commit-graph | 4.74 | 28.7 | **6.05×** | ±0.52 ms |
| `filtered-count-range-review-11` | 50,000 commits / 20,000 files; with commit-graph | 23.4 | 6.93 | 0.30× | ±1.07 ms |
| `filtered-count-range-review-6` | 50,000 commits / 20,000 files; with commit-graph | 3.26 | 5.19 | 1.59× | ±0.32 ms |
| `filtered-count-range-review-7` | 50,000 commits / 20,000 files; with commit-graph | 159 | 205 | 1.29× | ±2.46 ms |
| `filtered-count-range-review-8` | 50,000 commits / 20,000 files; with commit-graph | 4.73 | 29.1 | **6.16×** | ±1.33 ms |
| `filtered-count-range-review-9` | 50,000 commits / 20,000 files; with commit-graph | 2.88 | 5.08 | 1.76× | ±1.17 ms |
| `filtered-count-range-review-0` | 50,000 commits / 20,000 files; without commit-graph | 4.64 | 4.80 | 1.03× | ±0.22 ms |
| `filtered-count-range-review-1` | 50,000 commits / 20,000 files; without commit-graph | 152 | 204 | 1.35× | ±1.64 ms |
| `filtered-count-range-review-12` | 50,000 commits / 20,000 files; without commit-graph | 2.31 | 3.50 | 1.51× | ±0.28 ms |
| `filtered-count-range-review-13` | 50,000 commits / 20,000 files; without commit-graph | 3.80 | 8.18 | **2.15×** | ±0.28 ms |
| `filtered-count-range-review-2` | 50,000 commits / 20,000 files; without commit-graph | 4.25 | 197 | **46.28×** | ±2.16 ms |
| `filtered-count-range-review-3` | 50,000 commits / 20,000 files; without commit-graph | 2.81 | 4.58 | 1.63× | ±0.22 ms |
| `filtered-count-range-review-4` | 50,000 commits / 20,000 files; without commit-graph | 4.11 | 196 | **47.72×** | ±1.61 ms |
| `filtered-count-range-review-5` | 50,000 commits / 20,000 files; without commit-graph | 113 | 6.28 | 0.06× | ±0.66 ms |

### git-git-headlines

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `log/oneline-100@git-git` | git-git | 3.42 | 5.48 | 1.60× | ±0.59 ms |
| `resolve/head-n@git-git` | git-git | 1.32 | 3.21 | **2.43×** | ±0.23 ms |
| `show/head@git-git` | git-git | 2.04 | 3.81 | 1.86× | ±0.31 ms |
| `branch/list@many-refs` | many-refs | 1.45 | 8.76 | **6.03×** | ±0.75 ms |

### hash

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `index-pack-git-git` | /tmp/grit-bench-odb-cache/hot-path-100k-repacked | 196 | 123 | 0.63× | ±4.90 ms |
| `hash-object-256mib` | /tmp/grit-bench-scratch/hash-bench-256mib.bin | 6,237 | 3,810 | 0.61× | ±28.7 ms |

### history

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `blame-pack-2079-full` | git/git 7e34d8174dcda3435d1404c9a30bef46afe6c13e | 221 | 1,243 | **5.61×** | ±1.57 ms |
| `blame-pack-2079-range` | git/git 7e34d8174dcda3435d1404c9a30bef46afe6c13e | 172 | 436 | **2.53×** | ±4.97 ms |
| `history-drop-100` | history-rewrite-prod | 763 | 170 | 0.22× | ±2.40 ms |
| `history-fixup-500` | history-rewrite-prod | 153 | 561 | **3.68×** | ±7.61 ms |
| `history-reword-1000` | history-rewrite-prod | 6,617 | 188 | 0.03× | ±6.64 ms |
| `history-reword-1000-rebase` | history-rewrite-prod | 6,766 | 185 | 0.03× | ±4.72 ms |
| `history-reword-update-refs-head` | history-rewrite-prod | 6,599 | 198 | 0.03× | ±1.12 ms |
| `history-squash-10` | history-rewrite-prod | 40.7 | 34.8 | 0.86× | ±1.06 ms |
| `rebase/onto@synth-10k` | history-rewrite-prod | 46.9 | 43.2 | 0.92× | ±0.00 ms |

### ignore

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `ignore-normalization-dist-siblings-1000-after` | 1000 existing files across 10 directories; *.txt; empty index | 1.26 | 4.01 | **3.19×** | ±0.09 ms |
| `ignore-normalization-dist-siblings-1000-before` | 1000 existing files across 10 directories; *.txt; empty index | 1.26 | 5.46 | **4.34×** | ±0.16 ms |
| `ignore-normalization-release-siblings-1000-after` | 1000 existing files across 10 directories; *.txt; empty index | 1.28 | 4.35 | **3.39×** | ±0.21 ms |
| `ignore-normalization-release-siblings-1000-before` | 1000 existing files across 10 directories; *.txt; empty index | 1.28 | 5.88 | **4.58×** | ±0.41 ms |
| `ignore-normalization-dist-tracked-1000-after` | 1000 existing files across 10 directories; *.txt; git add -f . | 10.7 | 3.52 | 0.33× | ±0.08 ms |
| `ignore-normalization-dist-tracked-1000-before` | 1000 existing files across 10 directories; *.txt; git add -f . | 10.7 | 4.94 | 0.46× | ±0.12 ms |
| `ignore-normalization-release-tracked-1000-after` | 1000 existing files across 10 directories; *.txt; git add -f . | 10.5 | 3.70 | 0.35× | ±0.09 ms |
| `ignore-normalization-release-tracked-1000-before` | 1000 existing files across 10 directories; *.txt; git add -f . | 10.5 | 5.15 | 0.49× | ±0.10 ms |
| `ignore-normalization-dist-deep-1000-after` | 1000 existing files below 12 nested directories; *.txt; empty index | 1.69 | 9.62 | **5.69×** | ±0.32 ms |
| `ignore-normalization-dist-deep-1000-before` | 1000 existing files below 12 nested directories; *.txt; empty index | 1.69 | 11.5 | **6.80×** | ±0.37 ms |
| `ignore-normalization-release-deep-1000-after` | 1000 existing files below 12 nested directories; *.txt; empty index | 1.64 | 10.1 | **6.16×** | ±0.46 ms |
| `ignore-normalization-release-deep-1000-before` | 1000 existing files below 12 nested directories; *.txt; empty index | 1.64 | 11.8 | **7.22×** | ±0.19 ms |
| `ignore-normalization-dist-check-1000@synth-10k-after` | daily synth-10k; index emptied; *.txt; d0000/f0000.txt through d0999/f0999.txt | 2.25 | 5.79 | **2.58×** | ±0.17 ms |
| `ignore-normalization-dist-check-1000@synth-10k-before` | daily synth-10k; index emptied; *.txt; d0000/f0000.txt through d0999/f0999.txt | 2.25 | 7.63 | **3.40×** | ±0.19 ms |
| `ignore-normalization-release-check-1000@synth-10k-after` | daily synth-10k; index emptied; *.txt; d0000/f0000.txt through d0999/f0999.txt | 2.20 | 6.09 | **2.77×** | ±0.15 ms |
| `ignore-normalization-release-check-1000@synth-10k-before` | daily synth-10k; index emptied; *.txt; d0000/f0000.txt through d0999/f0999.txt | 2.20 | 8.01 | **3.64×** | ±0.23 ms |
| `ignore-normalization-dist-check-1@synth-10k-after` | first path from daily synth-10k ignore fixture | 1.04 | 1.55 | 1.50× | ±0.07 ms |
| `ignore-normalization-dist-check-1@synth-10k-before` | first path from daily synth-10k ignore fixture | 1.04 | 1.56 | 1.51× | ±0.06 ms |
| `ignore-normalization-release-check-1@synth-10k-after` | first path from daily synth-10k ignore fixture | 1.02 | 1.70 | 1.67× | ±0.06 ms |
| `ignore-normalization-release-check-1@synth-10k-before` | first path from daily synth-10k ignore fixture | 1.02 | 1.78 | 1.75× | ±0.07 ms |
| `ignore-check-1000` | synthetic-1000 | 1.91 | 3.51 | 1.84× | ±0.26 ms |

### large-blob-materialization

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `clone-small-5000-loose-1849` | 5000-small-loose-files | 504 | 424 | 0.84× | ±5.35 ms |
| `cat-loose-1849` | seq-40000000-loose | 737 | 540 | 0.73× | ±0.00 ms |
| `clone-loose-1849` | seq-40000000-loose | 880 | 503 | 0.57× | ±5.77 ms |
| `restore-loose-1849` | seq-40000000-loose | 890 | 760 | 0.85× | ±0.00 ms |
| `switch-loose-1849` | seq-40000000-loose | 877 | 500 | 0.57× | ±0.00 ms |

### log

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/log` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 5.37 | 0.00 | 0.00× | ±0.00 ms |
| `chromium-synth/log-patch` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 46,492 | 472,634 | **10.17×** | ±8,414 ms |
| `git-dot-git/log` | git-dot-git | 18.0 | 0.00 | 0.00× | ±0.00 ms |
| `git-dot-git/log-patch` | git-dot-git | 1,906 | 22,825 | **11.98×** | ±303 ms |
| `log-10000` | synthetic-10000 | 3.45 | 4.87 | 1.41× | ±0.43 ms |
| `log-100000` | synthetic-100000 | 3.43 | 4.75 | 1.38× | ±0.39 ms |

### maintenance

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `fsck-medium` | /tmp/grit-bench-network-cache/deep-history-prod.git | 193 | 159 | 0.83× | ±6.80 ms |
| `commit-graph-write` | /tmp/grit-bench-network-cache/deep-history-smoke.git | 3.88 | 4.54 | 1.17× | ±0.58 ms |
| `gc-medium` | /tmp/grit-bench-network-cache/deep-history-smoke.git | 22.4 | 6.83 | 0.30× | ±2.14 ms |
| `repack-medium` | /tmp/grit-bench-network-cache/deep-history-smoke.git | 12.8 | 21.2 | 1.66× | ±3.58 ms |
| `gc-loose-history` | /tmp/grit-bench-scratch/gc-loose-prod/history | 3,676 | 3,996 | 1.09× | ±63.8 ms |
| `gc-loose-wide` | /tmp/grit-bench-scratch/gc-loose-prod/wide | 7,313 | 6,909 | 0.94× | ±121 ms |
| `chromium-synth/fsck` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 2,020 | 78,713 | **38.97×** | ±391 ms |
| `chromium-synth/gc` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 2,351 | 0.00 | 0.00× | ±0.00 ms |
| `chromium-synth/repack` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 195,644 | 0.00 | 0.00× | ±0.00 ms |
| `git-dot-git/fsck` | git-dot-git | 22,665 | 0.00 | 0.00× | ±0.00 ms |
| `git-dot-git/gc` | git-dot-git | 4,352 | 0.00 | 0.00× | ±0.00 ms |
| `git-dot-git/repack` | git-dot-git | 24,684 | 0.00 | 0.00× | ±0.00 ms |

### merge

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `merge-content-10000-after` | 10000 | 1,911 | 1,483 | 0.78× | ±27.0 ms |
| `merge-content-10000-before` | 10000 | 1,911 | 1,519 | 0.79× | ±25.0 ms |
| `merge-content-20000-after` | 20000 | 3,813 | 2,931 | 0.77× | ±60.0 ms |
| `merge-content-20000-before` | 20000 | 3,813 | 3,004 | 0.79× | ±71.3 ms |
| `chromium-synth/merge` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 600 | 10,747 | **17.90×** | ±76.3 ms |
| `merge-content-daily-after` | daily | 2,469 | 2,708 | 1.10× | ±58.5 ms |
| `merge-content-daily-before` | daily | 2,469 | 2,812 | 1.14× | ±64.2 ms |
| `git-dot-git/merge` | git-dot-git | 12.8 | 83.0 | **6.48×** | ±2.59 ms |
| `merge-10000` | synthetic-10000 | 18.6 | 29.0 | 1.56× | ±1.70 ms |
| `merge-check` | synthetic-10000 | 1.49 | 92.2 | **62.06×** | ±0.80 ms |
| `merge-100000` | synthetic-100000 | 149 | 157 | 1.05× | ±4.96 ms |

### merge-base

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/merge-base` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 1.14 | 0.00 | 0.00× | ±0.00 ms |
| `git-dot-git/merge-base` | git-dot-git | 1.39 | 0.00 | 0.00× | ±0.00 ms |
| `merge-base-pick-fixture` | synthetic-10000 | 1.09 | 26.5 | **24.39×** | ±0.37 ms |

### midx

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `issue2064-midx_lookup-hit` | 10-pack synthetic fixture | — | 0.00 | — | ±0.00 ms |
| `issue2064-midx_lookup-miss` | 10-pack synthetic fixture | — | 0.00 | — | ±0.00 ms |
| `issue2064-midx_prepare-strict_and_permissive_cold` | git/git 422320 objects | — | 7.36 | — | ±0.04 ms |

### network

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/clone-file` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 15,342 | 729,454 | **47.55×** | ±2,464 ms |
| `chromium-synth/fetch-incr-file` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 1,629 | 3,226 | 1.98× | ±74.8 ms |
| `chromium-synth/push-file` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 1,233 | 2,727 | **2.21×** | ±129 ms |
| `git-dot-git/clone-file` | git-dot-git | 8,818 | 458 | 0.05× | ±3.90 ms |
| `git-dot-git/fetch-incr-file` | git-dot-git | 44.6 | 1,264 | **28.36×** | ±42.2 ms |
| `git-dot-git/push-file` | git-dot-git | 54.6 | 742 | **13.57×** | ±34.9 ms |

### network-clone

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `clone-file-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 548 | 517 | 0.94× | ±17.8 ms |
| `clone-git-http-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 567 | 1,049 | 1.85× | ±12.2 ms |
| `clone-grit-http-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 17,182 | 17,590 | 1.02× | ±29.4 ms |

### network-fetch

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `fetch-incr-file-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 26.8 | 136 | **5.10×** | ±6.38 ms |
| `fetch-incr-git-http-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 47.0 | 80.3 | 1.71× | ±1.15 ms |
| `fetch-incr-grit-http-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 279 | 624 | **2.24×** | ±18.6 ms |
| `fetch-noop-file-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 8.06 | 5.20 | 0.65× | ±0.26 ms |
| `fetch-noop-git-http-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 16.4 | 12.9 | 0.79× | ±0.89 ms |
| `fetch-noop-grit-http-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 11.8 | 10.2 | 0.87× | ±0.31 ms |
| `fetch-local-tags-noop-after` | 3000 linear commits, 1500 lightweight tags on even revisions + 1000 extra tip tags; packed file:// source | 51.6 | 25.9 | 0.50× | ±0.24 ms |
| `fetch-local-tags-noop-before` | 3000 linear commits, 1500 lightweight tags on even revisions + 1000 extra tip tags; packed file:// source | 51.6 | 26.5 | 0.51× | ±0.55 ms |
| `fetch-local-tags-noop-shallow-after` | 3000 linear commits, 1500 lightweight tags on even revisions + 1000 extra tip tags; packed file:// source | 26.8 | 32.7 | 1.22× | ±1.01 ms |
| `fetch-local-tags-noop-shallow-before` | 3000 linear commits, 1500 lightweight tags on even revisions + 1000 extra tip tags; packed file:// source | 26.8 | 32.9 | 1.23× | ±0.37 ms |
| `fetch-local-tags-unshallow-after` | 3000 linear commits, 1500 lightweight tags on even revisions + 1000 extra tip tags; packed file:// source | 229 | 492 | **2.15×** | ±3.00 ms |
| `fetch-local-tags-unshallow-before` | 3000 linear commits, 1500 lightweight tags on even revisions + 1000 extra tip tags; packed file:// source | 229 | 212 | 0.93× | ±1.94 ms |
| `fetch-local-tags-unshallow-complete-after` | 3000 linear commits, 1500 lightweight tags on even revisions + 1000 extra tip tags; packed file:// source | 227 | 496 | **2.19×** | ±4.86 ms |
| `fetch-local-tags-unshallow-complete-before` | 3000 linear commits, 1500 lightweight tags on even revisions + 1000 extra tip tags; packed file:// source | 227 | 538 | **2.37×** | ±8.05 ms |
| `fetch-local-deep-default-after` | git/git 82563 commits + 3000 heads / 2000 tags | 47.9 | 41.0 | 0.85× | ±0.23 ms |
| `fetch-local-deep-default-before` | git/git 82563 commits + 3000 heads / 2000 tags | 47.9 | 220 | **4.58×** | ±2.66 ms |
| `fetch-local-deep-no-tags-after` | git/git 82563 commits + 3000 heads / 2000 tags | 40.7 | 24.7 | 0.61× | ±0.38 ms |
| `fetch-local-deep-no-tags-before` | git/git 82563 commits + 3000 heads / 2000 tags | 40.7 | 204 | **5.00×** | ±1.78 ms |
| `clone-many-branches-200k-after` | one commit + 200,000 branches, packed (issue #2189) | 1,583 | 1,430 | 0.90× | ±26.5 ms |
| `clone-many-branches-200k-before` | one commit + 200,000 branches, packed (issue #2189) | 1,583 | 43,647 | **27.57×** | ±156 ms |
| `fetch-many-branches-200k-after` | one commit + 200,000 branches, packed (issue #2189) | 10,560 | 12,560 | 1.19× | ±306 ms |
| `fetch-many-branches-200k-before` | one commit + 200,000 branches, packed (issue #2189) | 10,560 | 42,353 | **4.01×** | ±685 ms |
| `clone-many-branches-25k-after` | one commit + 25,000 branches, packed (issue #2189) | 194 | 158 | 0.81× | ±13.0 ms |
| `clone-many-branches-25k-before` | one commit + 25,000 branches, packed (issue #2189) | 194 | 2,864 | **14.76×** | ±97.6 ms |
| `fetch-many-branches-25k-after` | one commit + 25,000 branches, packed (issue #2189) | 1,142 | 1,478 | 1.29× | ±51.2 ms |
| `fetch-many-branches-25k-before` | one commit + 25,000 branches, packed (issue #2189) | 1,142 | 2,758 | **2.42×** | ±90.4 ms |
| `fetch-local-single-default-after` | one commit + 3000 heads / 2000 tags | 41.3 | 35.5 | 0.86× | ±0.32 ms |
| `fetch-local-single-default-before` | one commit + 3000 heads / 2000 tags | 41.3 | 219 | **5.31×** | ±5.60 ms |
| `fetch-local-single-no-tags-after` | one commit + 3000 heads / 2000 tags | 37.2 | 24.1 | 0.65× | ±0.17 ms |
| `fetch-local-single-no-tags-before` | one commit + 3000 heads / 2000 tags | 37.2 | 211 | **5.66×** | ±4.76 ms |

### network-ls-remote

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `ls-remote-file-prod` | /tmp/grit-bench-network-cache/many-refs-prod.git | 10.5 | 16.7 | 1.60× | ±1.00 ms |
| `ls-remote-git-http-prod` | /tmp/grit-bench-network-cache/many-refs-prod.git | 20.0 | 20.0 | 1.00× | ±0.36 ms |
| `ls-remote-grit-http-prod` | /tmp/grit-bench-network-cache/many-refs-prod.git | 62.1 | 61.7 | 0.99× | ±0.81 ms |

### network-push

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `push-file-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 27.9 | 77.4 | **2.77×** | ±23.4 ms |
| `push-grit-http-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 28.8 | 171 | **5.93×** | ±5.53 ms |
| `push-one-commit-git-2a0f003c1` | git/git bare clone (82k commits, ~1,000 refs plus 137-257 pushed branches) | 90.6 | 1,840 | **20.30×** | ±65.3 ms |
| `push-one-commit-git-branch` | git/git bare clone (82k commits, ~1,000 refs plus 137-257 pushed branches) | 90.6 | 142 | 1.57× | ±19.8 ms |
| `push-one-commit-git-main` | git/git bare clone (82k commits, ~1,000 refs plus 137-257 pushed branches) | 90.6 | 499 | **5.50×** | ±56.3 ms |
| `push-one-commit-http-2a0f003c1` | git/git bare clone (82k commits, ~1,000 refs plus 137-257 pushed branches) | 48.7 | 1,798 | **36.93×** | ±59.0 ms |
| `push-one-commit-http-branch` | git/git bare clone (82k commits, ~1,000 refs plus 137-257 pushed branches) | 48.7 | 45.2 | 0.93× | ±1.73 ms |
| `push-one-commit-http-main` | git/git bare clone (82k commits, ~1,000 refs plus 137-257 pushed branches) | 48.7 | 454 | **9.32×** | ±53.6 ms |
| `push-one-commit-ssh-2a0f003c1` | git/git bare clone (82k commits, ~1,000 refs plus 137-257 pushed branches) | 64.0 | 1,794 | **28.02×** | ±65.0 ms |
| `push-one-commit-ssh-branch` | git/git bare clone (82k commits, ~1,000 refs plus 137-257 pushed branches) | 64.0 | 70.3 | 1.10× | ±4.50 ms |
| `push-one-commit-ssh-main` | git/git bare clone (82k commits, ~1,000 refs plus 137-257 pushed branches) | 64.0 | 453 | **7.07×** | ±60.7 ms |
| `push-one-commit-ssh-after` | git/git master 7e34d8174dcda3435d1404c9a30bef46afe6c13e (82563 commits) | 73.0 | 147 | **2.02×** | ±6.57 ms |
| `push-one-commit-ssh-before` | git/git master 7e34d8174dcda3435d1404c9a30bef46afe6c13e (82563 commits) | 73.0 | 2,631 | **36.03×** | ±22.1 ms |

### network-server

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `server-clone-http-compare-prod` | /tmp/grit-bench-network-cache/deep-history-prod.git | 556 | 17,061 | **30.67×** | ±38.3 ms |

### object_reads

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `cat-file-batch-sorted-git.git` | git.git | 32,433 | 46,558 | 1.44× | ±253 ms |
| `cat-file-batch-unordered-git.git` | git.git | 10,655 | 15,384 | 1.44× | ±62.4 ms |
| `log-patch-2000-git.git` | git.git | 1,518 | 699,640 | **460.95×** | ±15,497 ms |
| `rev-list-objects-git.git` | git.git | 3,036 | 1,590,307 | **523.82×** | ±2,433 ms |
| `cat-file-batch-sorted-hot-path-100k` | hot-path-100k | 119 | 922 | **7.76×** | ±2.48 ms |
| `cat-file-batch-unordered-hot-path-100k` | hot-path-100k | 117 | 937 | **8.03×** | ±28.4 ms |
| `log-patch-2000-hot-path-100k` | hot-path-100k | 400 | 1,827,664 | **4569.02×** | ±12,342 ms |
| `rev-list-objects-hot-path-100k` | hot-path-100k | 55.4 | 34,028 | **614.25×** | ±43.8 ms |

### odb

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `blame-pack-2079-cat-blob` | git/git 7e34d8174dcda3435d1404c9a30bef46afe6c13e | 2.89 | 3.70 | 1.28× | ±0.05 ms |
| `blame-pack-2079-cat-commit` | git/git 7e34d8174dcda3435d1404c9a30bef46afe6c13e | 2.33 | 3.24 | 1.39× | ±0.08 ms |
| `blame-pack-2079-cat-tree` | git/git 7e34d8174dcda3435d1404c9a30bef46afe6c13e | 2.60 | 8.91 | **3.43×** | ±0.06 ms |

### odb_backend

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `cat-file-batch-check-hot-path-100k` | hot-path-100k | 55.0 | 51.7 | 0.94× | ±7.14 ms |
| `cat-file-batch-check-hot-path-100k-pre-refactor` | hot-path-100k | 56.7 | 1,298 | **22.88×** | ±196 ms |
| `cat-file-batch-hot-path-100k` | hot-path-100k | 190 | 107 | 0.56× | ±15.9 ms |
| `cat-file-batch-hot-path-100k-pre-refactor` | hot-path-100k | 162 | 1,117 | **6.91×** | ±19.5 ms |
| `rev-list-objects-odb-backend-hot-path-100k` | hot-path-100k | 82.0 | 808 | **9.86×** | ±126 ms |
| `rev-list-objects-odb-backend-hot-path-100k-pre-refactor` | hot-path-100k | 71.9 | 7,062 | **98.20×** | ±446 ms |

### odb_read_parallel

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `odb-read-parallel-packed-blobs-1` | 50k-pack-blobs-3072B-loose-pruned | — | 149 | — | ±2.00 ms |
| `odb-read-parallel-packed-blobs-2` | 50k-pack-blobs-3072B-loose-pruned | — | 87.8 | — | ±1.50 ms |
| `odb-read-parallel-packed-blobs-4` | 50k-pack-blobs-3072B-loose-pruned | — | 50.9 | — | ±1.20 ms |
| `odb-read-parallel-packed-blobs-8` | 50k-pack-blobs-3072B-loose-pruned | — | 52.7 | — | ±1.00 ms |

### parallel

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `parallel-checkout-fresh-100000-t1` | synthetic-100000 | 1,235 | 2,444 | 1.98× | ±22.8 ms |
| `parallel-checkout-fresh-100000-tauto` | synthetic-100000 | 1,207 | 2,476 | **2.05×** | ±28.3 ms |
| `parallel-clone-file-100000-t1` | synthetic-100000 | 1,602 | 2,440 | 1.52× | ±18.7 ms |
| `parallel-clone-file-100000-tauto` | synthetic-100000 | 1,622 | 2,515 | 1.55× | ±85.7 ms |
| `parallel-diff-rename-100000-t1` | synthetic-100000 | 11.8 | 12.5 | 1.07× | ±0.55 ms |
| `parallel-diff-rename-100000-tauto` | synthetic-100000 | 11.6 | 12.2 | 1.05× | ±0.22 ms |
| `parallel-index-pack-100000-t1` | synthetic-100000 | 139 | 716 | **5.15×** | ±13.4 ms |
| `parallel-index-pack-100000-tauto` | synthetic-100000 | 166 | 338 | **2.03×** | ±3.45 ms |
| `parallel-pack-write-100000-t1` | synthetic-100000 | 103 | 27,233 | **264.04×** | ±59.7 ms |
| `parallel-pack-write-100000-tauto` | synthetic-100000 | 103 | 7,787 | **75.47×** | ±27.4 ms |
| `parallel-status-clean-100000-t1` | synthetic-100000 | 84.7 | 251 | **2.96×** | ±5.54 ms |
| `parallel-status-clean-100000-tauto` | synthetic-100000 | 65.8 | 214 | **3.25×** | ±8.99 ms |
| `parallel-status-dirty-100000-t1` | synthetic-100000 | 83.8 | 270 | **3.22×** | ±3.47 ms |
| `parallel-status-dirty-100000-tauto` | synthetic-100000 | 84.5 | 226 | **2.67×** | ±6.85 ms |
| `parallel-switch-wide-100000-t1` | synthetic-100000 | 370 | 4,828 | **13.06×** | ±62.6 ms |
| `parallel-switch-wide-100000-tauto` | synthetic-100000 | 351 | 4,723 | **13.46×** | ±41.2 ms |

### pick

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/pick` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 801 | 1,169,969 | **1461.52×** | ±32,222 ms |
| `git-dot-git/pick` | git-dot-git | 13.9 | 387 | **27.85×** | ±13.7 ms |
| `pick-10000` | synthetic-10000 | 114 | 200 | 1.75× | ±2.26 ms |
| `pick-range-20` | synthetic-10000 | 118 | 94.5 | 0.80× | ±4.21 ms |
| `pick-series-10000` | synthetic-10000 | 116 | 484 | **4.16×** | ±2.08 ms |
| `revert-10000` | synthetic-10000 | 106 | 199 | 1.88× | ±4.47 ms |
| `pick-100000` | synthetic-100000 | 196 | 1,796 | **9.17×** | ±30.8 ms |
| `pick-series-100000` | synthetic-100000 | 880 | 5,176 | **5.88×** | ±104 ms |
| `revert-100000` | synthetic-100000 | 199 | 1,820 | **9.14×** | ±39.7 ms |

### primary-store-7286

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `primary-store-7286-revs-count-synth-120k-after` | 120,000-commit linear history from git fast-import, one file changed per commit, repacked, no commit-graph | 392 | 262 | 0.67× | ±18.9 ms |
| `primary-store-7286-revs-count-synth-120k-before` | 120,000-commit linear history from git fast-import, one file changed per commit, repacked, no commit-graph | 392 | 270 | 0.69× | ±14.3 ms |
| `primary-store-7286-blame-file-git-git-after` | git.git worktree at the pinned grit-bench tag | 296 | 255 | 0.86× | ±12.8 ms |
| `primary-store-7286-blame-file-git-git-before` | git.git worktree at the pinned grit-bench tag | 296 | 258 | 0.87× | ±14.8 ms |
| `primary-store-7286-log-oneline-100-git-git-after` | git.git worktree at the pinned grit-bench tag | 3.27 | 4.04 | 1.24× | ±2.13 ms |
| `primary-store-7286-log-oneline-100-git-git-before` | git.git worktree at the pinned grit-bench tag | 3.27 | 3.95 | 1.21× | ±0.33 ms |
| `primary-store-7286-revs-count-git-git-after` | git.git worktree at the pinned grit-bench tag | 532 | 387 | 0.73× | ±17.1 ms |
| `primary-store-7286-revs-count-git-git-before` | git.git worktree at the pinned grit-bench tag | 532 | 403 | 0.76× | ±36.0 ms |
| `primary-store-7286-show-head-git-git-after` | git.git worktree at the pinned grit-bench tag | 2.24 | 2.71 | 1.21× | ±0.15 ms |
| `primary-store-7286-show-head-git-git-before` | git.git worktree at the pinned grit-bench tag | 2.24 | 2.71 | 1.21× | ±0.14 ms |
| `primary-store-7286-diff-worktree-synth-10k-after` | grit-bench synth-10k worktree with the diff/worktree edits applied | 6.96 | 6.81 | 0.98× | ±1.74 ms |
| `primary-store-7286-diff-worktree-synth-10k-before` | grit-bench synth-10k worktree with the diff/worktree edits applied | 6.96 | 6.62 | 0.95× | ±0.44 ms |
| `primary-store-7286-status-dirty-synth-10k-after` | grit-bench synth-10k worktree with the diff/worktree edits applied | 7.72 | 7.30 | 0.95× | ±0.41 ms |
| `primary-store-7286-status-dirty-synth-10k-before` | grit-bench synth-10k worktree with the diff/worktree edits applied | 7.72 | 7.27 | 0.94× | ±1.09 ms |

### ref-query

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `ref-query-contains-tags@git-git-no-graph` | git-git | 6,407 | 208,000 | **32.46×** | ±0.00 ms |
| `ref-query-contains-tags@git-git-with-graph` | git-git | 6,407 | 301 | 0.05× | ±0.00 ms |
| `ref-query-merged-branches@refs-100k-packed` | refs-100k-packed | 48.1 | 689 | **14.33×** | ±0.00 ms |

### refs

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `ref/list@100k-4184` | 100k | 58.6 | 128 | **2.19×** | ±3.66 ms |
| `tag/list@100k-4184` | 100k | 27.5 | 31.3 | 1.14× | ±0.84 ms |
| `branch/list@100k-local-4184` | 100k-local | 353 | 151 | 0.43× | ±6.50 ms |
| `branch/list@local-6667-4184` | local-6667 | 22.7 | 11.3 | 0.50× | ±0.24 ms |
| `branch/list@local-loose-6667-4184` | local-loose-6667 | 86.5 | 140 | 1.62× | ±0.69 ms |
| `branch/create@many-refs-4184` | many-refs | 1.57 | 2.72 | 1.73× | ±0.13 ms |
| `branch/delete@many-refs-4184` | many-refs | 1.85 | 8.00 | **4.32×** | ±4.57 ms |
| `branch/list-json@many-refs-4184` | many-refs | 1.65 | 2.35 | 1.43× | ±0.05 ms |
| `branch/list@many-refs-4184` | many-refs | 1.41 | 2.13 | 1.52× | ±0.06 ms |
| `branch/merged-ref@many-refs-4184` | many-refs | 1.56 | 2.83 | 1.82× | ±0.13 ms |
| `branch/merged@many-refs-4184` | many-refs | 1.60 | 2.45 | 1.53× | ±0.07 ms |
| `branch/rename@many-refs-4184` | many-refs | 1.84 | 7.10 | **3.85×** | ±0.27 ms |
| `ref/list@many-refs-4184` | many-refs | 24.5 | 27.3 | 1.11× | ±1.73 ms |
| `tag/create@many-refs-4184` | many-refs | 1.53 | 2.66 | 1.74× | ±0.11 ms |
| `tag/list@many-refs-4184` | many-refs | 4.05 | 5.10 | 1.26× | ±0.86 ms |
| `branch/list@many-refs-loose-4184` | many-refs-loose | 1.39 | 2.15 | 1.54× | ±0.08 ms |
| `for-each-ref-all-refs-100k-loose` | refs-100k-loose | 1,416 | 1,095 | 0.77× | ±48.4 ms |
| `for-each-ref-prefix-b05-refs-100k-loose` | refs-100k-loose | 436 | 160 | 0.37× | ±17.9 ms |
| `update-ref-create-100k-refs-100k-loose` | refs-100k-loose | 4,154 | 3,283 | 0.79× | ±136 ms |
| `update-ref-delete-1k-refs-100k-loose` | refs-100k-loose | 54.6 | 61.4 | 1.12× | ±16.3 ms |
| `update-ref-update-1k-refs-100k-loose` | refs-100k-loose | 75.2 | 71.9 | 0.96× | ±15.8 ms |
| `upload-pack-advertise-v0-refs-100k-loose` | refs-100k-loose | 301 | 1,234 | **4.09×** | ±53.3 ms |
| `upload-pack-ls-refs-prefix-refs-100k-loose` | refs-100k-loose | 275 | 42.9 | 0.16× | ±1.91 ms |
| `for-each-ref-all-refs-100k-packed` | refs-100k-packed | 987 | 693 | 0.70× | ±8.91 ms |
| `for-each-ref-prefix-b05-refs-100k-packed` | refs-100k-packed | 11.5 | 39.5 | **3.44×** | ±6.13 ms |
| `update-ref-create-100k-refs-100k-packed` | refs-100k-packed | 4,138 | 3,858 | 0.93× | ±227 ms |
| `update-ref-delete-1k-refs-100k-packed` | refs-100k-packed | 68.3 | 43.9 | 0.64× | ±20.0 ms |
| `update-ref-update-1k-refs-100k-packed` | refs-100k-packed | 62.7 | 58.5 | 0.93× | ±16.7 ms |
| `upload-pack-advertise-v0-refs-100k-packed` | refs-100k-packed | 16.4 | 118 | **7.22×** | ±4.75 ms |
| `upload-pack-ls-refs-prefix-refs-100k-packed` | refs-100k-packed | 1.07 | 4.10 | **3.85×** | ±0.51 ms |
| `for-each-ref-all-refs-100k-reftable` | refs-100k-reftable | 884 | 57.3 | 0.06× | ±5.39 ms |
| `for-each-ref-prefix-b05-refs-100k-reftable` | refs-100k-reftable | 10.7 | 9.35 | 0.87× | ±0.97 ms |
| `update-ref-create-100k-refs-100k-reftable` | refs-100k-reftable | 873 | 479 | 0.55× | ±28.2 ms |
| `update-ref-delete-1k-refs-100k-reftable` | refs-100k-reftable | 11.3 | 8.33 | 0.74× | ±0.73 ms |
| `update-ref-update-1k-refs-100k-reftable` | refs-100k-reftable | 23.2 | 8.89 | 0.38× | ±1.09 ms |

### reset

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `reset-blobless-cargo` | cargo | 679 | 757 | 1.12× | ±0.00 ms |
| `reset-clean-cargo` | cargo | 14.0 | 17.3 | 1.24× | ±0.32 ms |
| `reset-full-cargo` | cargo | 152 | 179 | 1.18× | ±1.43 ms |
| `reset-blobless-http-200` | http-200 | 162 | 69.6 | 0.43× | ±0.00 ms |
| `reset-blobless-small` | small | 47.4 | 33.0 | 0.70× | ±0.28 ms |
| `reset-clean-small` | small | 5.88 | 6.02 | 1.02× | ±0.07 ms |
| `reset-full-small` | small | 5.47 | 9.72 | 1.78× | ±0.45 ms |

### resolve

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/resolve` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 0.89 | 15.5 | **17.34×** | ±1.96 ms |
| `git-dot-git/resolve` | git-dot-git | 0.96 | 16.1 | **16.67×** | ±4.83 ms |

### resolve-cat

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `cat-blob` | synthetic-500-history-80 | 1.07 | 2.52 | **2.36×** | ±0.23 ms |
| `resolve-HEAD~20` | synthetic-500-history-80 | 1.35 | 3.05 | **2.27×** | ±0.31 ms |

### restore

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `large-blob-memory-restore-1024m-loose` | incompressible-1024m-loose | 760 | 953 | 1.25× | ±23.1 ms |
| `large-blob-memory-restore-1024m-packed` | incompressible-1024m-packed | 713 | 947 | 1.33× | ±11.5 ms |

### revert

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/revert` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 829 | 1,265,597 | **1527.28×** | ±36,115 ms |
| `git-dot-git/revert` | git-dot-git | 13.0 | 362 | **27.90×** | ±5.92 ms |

### revision-selection

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `log-author-n50-commit-graph` | /tmp/grit-bench-scratch/revision-selection-worktree-commit-graph | 1.70 | 3.39 | 2.00× | ±0.20 ms |
| `log-author-n50-no-graph` | /tmp/grit-bench-scratch/revision-selection-worktree-no-graph | 1.41 | 3.53 | **2.49×** | ±0.18 ms |

### revs

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `revs-author-n50-commit-graph` | /tmp/grit-bench-network-cache/deep-history-prod.git | 1.59 | 2.98 | 1.87× | ±0.69 ms |
| `revs-author-n50-no-graph` | /tmp/grit-bench-network-cache/deep-history-prod.git | 1.29 | 2.93 | **2.28×** | ±0.17 ms |
| `revs-count-author-commit-graph` | /tmp/grit-bench-network-cache/deep-history-prod.git | 119 | 234 | 1.97× | ±8.92 ms |
| `revs-count-author-no-graph` | /tmp/grit-bench-network-cache/deep-history-prod.git | 110 | 379 | **3.46×** | ±2.30 ms |
| `revs-count-deep-commit-graph` | /tmp/grit-bench-network-cache/deep-history-prod.git | 19.9 | 6.32 | 0.32× | ±0.65 ms |
| `revs-count-deep-no-graph` | /tmp/grit-bench-network-cache/deep-history-prod.git | 85.5 | 6.22 | 0.07× | ±0.14 ms |
| `revs-count-since-no-graph` | /tmp/grit-bench-network-cache/deep-history-prod.git | 1.21 | 2.72 | **2.24×** | ±0.18 ms |
| `revs-left-right-count-no-graph` | /tmp/grit-bench-network-cache/deep-history-prod.git | 2.13 | 441 | **206.68×** | ±5.63 ms |
| `revs-list-deep-commit-graph` | /tmp/grit-bench-network-cache/deep-history-prod.git | 29.9 | 72.8 | **2.43×** | ±7.00 ms |
| `revs-list-deep-no-graph` | /tmp/grit-bench-network-cache/deep-history-prod.git | 97.4 | 274 | **2.81×** | ±7.02 ms |

### revs-startup-4992

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `revs-startup-4992-log-author-n50-deep-history-cg-after` | 50,000 commits, 20,000 files; commit-graph written | 1.78 | 3.19 | 1.79× | ±0.16 ms |
| `revs-startup-4992-log-author-n50-deep-history-cg-before` | 50,000 commits, 20,000 files; commit-graph written | 1.78 | 3.21 | 1.80× | ±0.16 ms |
| `revs-startup-4992-revs-author-n50-deep-history-cg-after` | 50,000 commits, 20,000 files; commit-graph written | 1.77 | 2.97 | 1.67× | ±0.20 ms |
| `revs-startup-4992-revs-author-n50-deep-history-cg-before` | 50,000 commits, 20,000 files; commit-graph written | 1.77 | 3.18 | 1.80× | ±0.22 ms |
| `revs-startup-4992-revs-count-author-deep-history-cg-after` | 50,000 commits, 20,000 files; commit-graph written | 124 | 130 | 1.05× | ±16.8 ms |
| `revs-startup-4992-revs-count-author-deep-history-cg-before` | 50,000 commits, 20,000 files; commit-graph written | 124 | 125 | 1.01× | ±5.36 ms |
| `revs-startup-4992-log-author-n50-deep-history-after` | 50,000 commits, 20,000 files; no commit-graph | 1.52 | 3.06 | **2.01×** | ±0.15 ms |
| `revs-startup-4992-log-author-n50-deep-history-before` | 50,000 commits, 20,000 files; no commit-graph | 1.52 | 3.08 | **2.02×** | ±0.16 ms |
| `revs-startup-4992-revs-author-n50-deep-history-after` | 50,000 commits, 20,000 files; no commit-graph | 1.50 | 2.89 | 1.92× | ±0.20 ms |
| `revs-startup-4992-revs-author-n50-deep-history-before` | 50,000 commits, 20,000 files; no commit-graph | 1.50 | 3.10 | **2.06×** | ±0.21 ms |
| `revs-startup-4992-revs-count-author-deep-history-after` | 50,000 commits, 20,000 files; no commit-graph | 114 | 131 | 1.15× | ±15.5 ms |
| `revs-startup-4992-revs-count-author-deep-history-before` | 50,000 commits, 20,000 files; no commit-graph | 114 | 126 | 1.11× | ±5.50 ms |
| `revs-startup-4992-revs-json-author-n50-deep-history-after` | 50,000 commits, 20,000 files; no commit-graph | 1.46 | 3.06 | **2.10×** | ±0.26 ms |
| `revs-startup-4992-revs-json-author-n50-deep-history-before` | 50,000 commits, 20,000 files; no commit-graph | 1.46 | 3.06 | **2.10×** | ±0.26 ms |
| `revs-startup-4992-revs-left-right-count-deep-history-after` | 50,000 commits, 20,000 files; no commit-graph | 2.19 | 4.01 | 1.83× | ±0.30 ms |
| `revs-startup-4992-revs-left-right-count-deep-history-before` | 50,000 commits, 20,000 files; no commit-graph | 2.19 | 4.03 | 1.84× | ±0.25 ms |
| `revs-startup-4992-log-json-100-git-git-after` | git/git v2.47.0 | 2.81 | 4.42 | 1.57× | ±0.35 ms |
| `revs-startup-4992-log-json-100-git-git-before` | git/git v2.47.0 | 2.81 | 4.44 | 1.58× | ±0.30 ms |
| `revs-startup-4992-log-oneline-100-git-git-after` | git/git v2.47.0 | 2.88 | 4.42 | 1.54× | ±0.34 ms |
| `revs-startup-4992-log-oneline-100-git-git-before` | git/git v2.47.0 | 2.88 | 4.50 | 1.57× | ±0.43 ms |
| `revs-startup-4992-revs-count-git-git-after` | git/git v2.47.0 | 520 | 670 | 1.29× | ±16.9 ms |
| `revs-startup-4992-revs-count-git-git-before` | git/git v2.47.0 | 520 | 667 | 1.28× | ±22.4 ms |
| `revs-startup-4992-show-head-git-git-after` | git/git v2.47.0 | 1.73 | 2.75 | 1.59× | ±0.25 ms |
| `revs-startup-4992-show-head-git-git-before` | git/git v2.47.0 | 1.73 | 2.74 | 1.59× | ±0.25 ms |

### startup-relocations

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `startup-dist-branch@deep` | deep | 1.21 | 1.98 | 1.64× | ±0.10 ms |
| `startup-dist-help@deep` | deep | 0.80 | 2.79 | **3.47×** | ±0.11 ms |
| `startup-dist-log@deep` | deep | 6.67 | 9.57 | 1.43× | ±0.19 ms |
| `startup-dist-resolve@deep` | deep | 0.92 | 1.84 | **2.01×** | ±0.11 ms |
| `startup-dist-status@deep` | deep | 1.37 | 2.31 | 1.69× | ±0.09 ms |
| `startup-release-branch@deep` | deep | 1.33 | 2.32 | 1.75× | ±0.06 ms |
| `startup-release-help@deep` | deep | 0.85 | 2.90 | **3.39×** | ±0.09 ms |
| `startup-release-log@deep` | deep | 6.33 | 9.21 | 1.46× | ±0.19 ms |
| `startup-release-resolve@deep` | deep | 1.00 | 2.19 | **2.19×** | ±0.08 ms |
| `startup-release-status@deep` | deep | 1.66 | 2.92 | 1.77× | ±0.09 ms |
| `startup-review-dist-branch@deep` | deep | 1.43 | 2.24 | 1.57× | ±0.06 ms |
| `startup-review-dist-help@deep` | deep | 0.93 | 3.00 | **3.21×** | ±0.08 ms |
| `startup-review-dist-log@deep` | deep | 7.11 | 10.1 | 1.41× | ±0.13 ms |
| `startup-review-dist-resolve@deep` | deep | 1.07 | 2.05 | 1.91× | ±0.03 ms |
| `startup-review-dist-status@deep` | deep | 1.77 | 2.73 | 1.54× | ±0.05 ms |
| `startup-review-release-branch@deep` | deep | 1.46 | 2.83 | 1.94× | ±0.07 ms |
| `startup-review-release-help@deep` | deep | 0.94 | 3.34 | **3.57×** | ±0.13 ms |
| `startup-review-release-log@deep` | deep | 7.14 | 11.3 | 1.58× | ±0.26 ms |
| `startup-review-release-resolve@deep` | deep | 1.09 | 2.70 | **2.47×** | ±0.08 ms |
| `startup-review-release-status@deep` | deep | 1.79 | 3.49 | 1.95× | ±0.11 ms |
| `startup-dist-branch@many-refs` | many-refs | 45.2 | 11.5 | 0.25× | ±0.37 ms |
| `startup-dist-help@many-refs` | many-refs | 0.79 | 2.73 | **3.47×** | ±0.15 ms |
| `startup-dist-log@many-refs` | many-refs | 2.53 | 2.63 | 1.04× | ±0.09 ms |
| `startup-dist-resolve@many-refs` | many-refs | 0.91 | 1.82 | 2.00× | ±0.07 ms |
| `startup-dist-status@many-refs` | many-refs | 1.37 | 2.30 | 1.67× | ±0.08 ms |
| `startup-release-branch@many-refs` | many-refs | 43.9 | 10.9 | 0.25× | ±0.41 ms |
| `startup-release-help@many-refs` | many-refs | 0.89 | 3.01 | **3.40×** | ±0.14 ms |
| `startup-release-log@many-refs` | many-refs | 2.88 | 3.02 | 1.05× | ±0.09 ms |
| `startup-release-resolve@many-refs` | many-refs | 0.99 | 2.14 | **2.16×** | ±0.09 ms |
| `startup-release-status@many-refs` | many-refs | 1.56 | 2.78 | 1.78× | ±0.10 ms |
| `startup-review-dist-branch@many-refs` | many-refs | 51.4 | 13.5 | 0.26× | ±1.22 ms |
| `startup-review-dist-help@many-refs` | many-refs | 0.93 | 3.03 | **3.24×** | ±0.09 ms |
| `startup-review-dist-log@many-refs` | many-refs | 3.17 | 2.84 | 0.90× | ±0.06 ms |
| `startup-review-dist-resolve@many-refs` | many-refs | 1.10 | 2.09 | 1.90× | ±0.04 ms |
| `startup-review-dist-status@many-refs` | many-refs | 1.78 | 2.75 | 1.55× | ±0.05 ms |
| `startup-review-release-branch@many-refs` | many-refs | 51.6 | 14.1 | 0.27× | ±0.37 ms |
| `startup-review-release-help@many-refs` | many-refs | 0.94 | 3.36 | **3.59×** | ±0.06 ms |
| `startup-review-release-log@many-refs` | many-refs | 3.20 | 3.61 | 1.13× | ±0.09 ms |
| `startup-review-release-resolve@many-refs` | many-refs | 1.10 | 2.69 | **2.45×** | ±0.08 ms |
| `startup-review-release-status@many-refs` | many-refs | 1.79 | 3.48 | 1.94× | ±0.09 ms |
| `startup-dist-branch@tiny` | tiny | 0.99 | 1.75 | 1.76× | ±0.08 ms |
| `startup-dist-help@tiny` | tiny | 0.81 | 2.77 | **3.43×** | ±0.11 ms |
| `startup-dist-log@tiny` | tiny | 1.12 | 1.95 | 1.75× | ±0.12 ms |
| `startup-dist-resolve@tiny` | tiny | 0.92 | 1.79 | 1.96× | ±0.07 ms |
| `startup-dist-status@tiny` | tiny | 1.26 | 2.14 | 1.69× | ±0.08 ms |
| `startup-release-branch@tiny` | tiny | 1.14 | 2.09 | 1.84× | ±0.08 ms |
| `startup-release-help@tiny` | tiny | 0.90 | 3.04 | **3.39×** | ±0.12 ms |
| `startup-release-log@tiny` | tiny | 1.28 | 2.33 | 1.83× | ±0.09 ms |
| `startup-release-resolve@tiny` | tiny | 1.04 | 2.24 | **2.15×** | ±0.09 ms |
| `startup-release-status@tiny` | tiny | 1.47 | 2.68 | 1.82× | ±0.14 ms |
| `startup-review-dist-branch@tiny` | tiny | 1.20 | 2.00 | 1.66× | ±0.13 ms |
| `startup-review-dist-help@tiny` | tiny | 0.94 | 3.05 | **3.25×** | ±0.12 ms |
| `startup-review-dist-log@tiny` | tiny | 1.27 | 2.09 | 1.64× | ±0.18 ms |
| `startup-review-dist-resolve@tiny` | tiny | 1.06 | 2.01 | 1.89× | ±0.04 ms |
| `startup-review-dist-status@tiny` | tiny | 1.51 | 2.38 | 1.57× | ±0.04 ms |
| `startup-review-release-branch@tiny` | tiny | 1.20 | 2.53 | **2.10×** | ±0.14 ms |
| `startup-review-release-help@tiny` | tiny | 0.94 | 3.35 | **3.54×** | ±0.12 ms |
| `startup-review-release-log@tiny` | tiny | 1.29 | 2.72 | **2.10×** | ±0.07 ms |
| `startup-review-release-resolve@tiny` | tiny | 1.08 | 2.62 | **2.42×** | ±0.09 ms |
| `startup-review-release-status@tiny` | tiny | 1.53 | 3.12 | **2.04×** | ±0.09 ms |
| `startup-dist-branch@wide` | wide | 1.21 | 1.95 | 1.61× | ±0.06 ms |
| `startup-dist-help@wide` | wide | 0.80 | 2.82 | **3.51×** | ±0.18 ms |
| `startup-dist-log@wide` | wide | 1.73 | 2.65 | 1.53× | ±0.09 ms |
| `startup-dist-resolve@wide` | wide | 0.92 | 1.82 | 1.98× | ±0.07 ms |
| `startup-dist-status@wide` | wide | 12.8 | 13.4 | 1.05× | ±0.51 ms |
| `startup-release-branch@wide` | wide | 1.38 | 2.37 | 1.72× | ±0.09 ms |
| `startup-release-help@wide` | wide | 0.90 | 3.02 | **3.38×** | ±0.12 ms |
| `startup-release-log@wide` | wide | 1.92 | 3.03 | 1.57× | ±0.09 ms |
| `startup-release-resolve@wide` | wide | 1.06 | 2.25 | **2.13×** | ±0.11 ms |
| `startup-release-status@wide` | wide | 16.2 | 16.5 | 1.02× | ±0.40 ms |
| `startup-review-dist-branch@wide` | wide | 1.42 | 2.22 | 1.56× | ±0.04 ms |
| `startup-review-dist-help@wide` | wide | 0.93 | 3.01 | **3.22×** | ±0.09 ms |
| `startup-review-dist-log@wide` | wide | 2.03 | 2.85 | 1.40× | ±0.04 ms |
| `startup-review-dist-resolve@wide` | wide | 1.07 | 2.05 | 1.91× | ±0.03 ms |
| `startup-review-dist-status@wide` | wide | 16.3 | 16.7 | 1.02× | ±0.28 ms |
| `startup-review-release-branch@wide` | wide | 1.44 | 2.84 | 1.97× | ±0.20 ms |
| `startup-review-release-help@wide` | wide | 0.94 | 3.36 | **3.59×** | ±0.09 ms |
| `startup-review-release-log@wide` | wide | 2.03 | 3.58 | 1.76× | ±0.08 ms |
| `startup-review-release-resolve@wide` | wide | 1.08 | 2.68 | **2.48×** | ±0.07 ms |
| `startup-review-release-status@wide` | wide | 16.2 | 18.2 | 1.12× | ±1.36 ms |

### stash

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/stash-push-pop` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 9,924 | 183,778 | **18.52×** | ±2,593 ms |
| `git-dot-git/stash-push-pop` | git-dot-git | 128 | 0.00 | 0.00× | ±0.00 ms |
| `stash-push-pop-10000` | synthetic-10000 | 358 | 518 | 1.44× | ±2.36 ms |
| `stash-push-pop-100000` | synthetic-100000 | 3,517 | 13,573 | **3.86×** | ±360 ms |

### status

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/status-clean` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 429 | 1,181 | **2.75×** | ±51.5 ms |
| `chromium-synth/status-dirty` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 574 | 1,222 | **2.13×** | ±28.1 ms |
| `git-dot-git/status-clean` | git-dot-git | 6.57 | 29.3 | **4.46×** | ±6.61 ms |
| `git-dot-git/status-dirty` | git-dot-git | 7.65 | 45.8 | **5.99×** | ±13.1 ms |
| `status-midx-2064-human-absent` | git/git 7e34d817; packed refs, changed-path commit graph; MIDX absent | 13.3 | 15.9 | 1.20× | ±0.13 ms |
| `status-midx-2064-json-absent` | git/git 7e34d817; packed refs, changed-path commit graph; MIDX absent | 13.5 | 15.9 | 1.18× | ±0.13 ms |
| `status-midx-2064-markdown-absent` | git/git 7e34d817; packed refs, changed-path commit graph; MIDX absent | 13.2 | 15.8 | 1.20× | ±0.11 ms |
| `status-midx-2064-short-absent` | git/git 7e34d817; packed refs, changed-path commit graph; MIDX absent | 13.3 | 15.9 | 1.20× | ±0.25 ms |
| `status-midx-2064-human-present` | git/git 7e34d817; packed refs, changed-path commit graph; MIDX present | 14.6 | 24.8 | 1.71× | ±1.15 ms |
| `status-midx-2064-json-present` | git/git 7e34d817; packed refs, changed-path commit graph; MIDX present | 13.6 | 24.3 | 1.78× | ±0.41 ms |
| `status-midx-2064-markdown-present` | git/git 7e34d817; packed refs, changed-path commit graph; MIDX present | 13.8 | 24.2 | 1.75× | ±0.40 ms |
| `status-midx-2064-short-present` | git/git 7e34d817; packed refs, changed-path commit graph; MIDX present | 13.5 | 24.2 | 1.79× | ±0.34 ms |
| `large-blob-memory-status-1024m` | incompressible-1024m | 2,560 | 1,030 | 0.40× | ±0.00 ms |
| `large-blob-memory-status-128m` | incompressible-128m | 360 | 90.0 | 0.25× | ±0.00 ms |
| `status/clean@small-4184` | small | 1.92 | 2.83 | 1.48× | ±0.11 ms |
| `status-clean-10000` | synthetic-10000 | 9.02 | 8.93 | 0.99× | ±4.13 ms |
| `status-dirty-10000` | synthetic-10000 | 10.9 | 10.5 | 0.96× | ±4.75 ms |
| `status-clean-100000` | synthetic-100000 | 78.0 | 83.6 | 1.07× | ±5.51 ms |
| `status-dirty-100000` | synthetic-100000 | 100 | 91.2 | 0.91× | ±5.32 ms |

### switch

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/switch` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 2,136 | 28,878 | **13.52×** | ±394 ms |
| `chromium-synth/switch-wide` | chromium-synth-f400000-d40000-c100000-r50000-s1129466183 | 2,398 | 28,414 | **11.85×** | ±534 ms |
| `git-dot-git/switch` | git-dot-git | 33.4 | 191 | **5.71×** | ±6.31 ms |
| `git-dot-git/switch-wide` | git-dot-git | 31.2 | 187 | **6.01×** | ±3.19 ms |
| `switch-10000` | synthetic-10000 | 22.0 | 27.5 | 1.25× | ±1.42 ms |
| `switch-wide-10000` | synthetic-10000 | 46.9 | 50.1 | 1.07× | ±2.19 ms |
| `switch-100000` | synthetic-100000 | 155 | 103 | 0.66× | ±4.31 ms |
| `switch-wide-100000` | synthetic-100000 | 380 | 359 | 0.95× | ±13.0 ms |

### tag-tree-sha256

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `chromium-synth/tag-sign-tree-sha256` | chromium-synth-f200-d20-c50-r10-s42 | 32.1 | 30.8 | 0.96× | ±2.68 ms |
| `chromium-synth/tag-verify-digest-no-signature` | chromium-synth-f200-d20-c50-r10-s42 | 31.7 | 4.34 | 0.14× | ±0.80 ms |
| `chromium-synth/tag-verify-signed` | chromium-synth-f200-d20-c50-r10-s42 | 8.71 | 10.2 | 1.17× | ±0.87 ms |
| `git-dot-git/tag-sign-tree-sha256` | git-dot-git | 28.1 | 134 | **4.75×** | ±5.80 ms |
| `git-dot-git/tag-verify-digest-no-signature` | git-dot-git | 388 | 134 | 0.35× | ±32.2 ms |
| `git-dot-git/tag-verify-signed` | git-dot-git | 8.98 | 118 | **13.15×** | ±11.4 ms |

### tree-metadata

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `tree-metadata-cli-2083-loose-api` | 120000 files, 400 directories, 20000 modified paths; loose | 50.5 | 57.6 | 1.14× | ±0.54 ms |
| `tree-metadata-cli-2083-loose-cli-name-only` | 120000 files, 400 directories, 20000 modified paths; loose | 50.9 | 52.7 | 1.04× | ±0.51 ms |
| `tree-metadata-cli-2083-loose-cli-name-only-dist` | 120000 files, 400 directories, 20000 modified paths; loose | 50.9 | 50.6 | 1.00× | ±0.92 ms |
| `tree-metadata-cli-2083-loose-cli-name-status` | 120000 files, 400 directories, 20000 modified paths; loose | 50.5 | 52.9 | 1.05× | ±1.61 ms |
| `tree-metadata-cli-2083-loose-cli-name-status-dist` | 120000 files, 400 directories, 20000 modified paths; loose | 50.5 | 50.4 | 1.00× | ±0.52 ms |
| `tree-metadata-cli-2083-loose-cli-name-status-json` | 120000 files, 400 directories, 20000 modified paths; loose | 50.5 | 56.9 | 1.13× | ±2.08 ms |
| `tree-metadata-cli-2083-loose-cli-name-status-json-dist` | 120000 files, 400 directories, 20000 modified paths; loose | 50.5 | 55.0 | 1.09× | ±2.95 ms |
| `tree-metadata-cli-2083-loose-cli-name-status-no-renames` | 120000 files, 400 directories, 20000 modified paths; loose | 50.6 | 53.9 | 1.06× | ±1.08 ms |
| `tree-metadata-cli-2083-loose-cli-name-status-no-renames-dist` | 120000 files, 400 directories, 20000 modified paths; loose | 50.6 | 51.6 | 1.02× | ±2.67 ms |
| `tree-metadata-loose-api` | 120000 files, 400 directories, 20000 modified paths; loose | 37.8 | 58.5 | 1.55× | ±0.64 ms |
| `tree-metadata-loose-cli` | 120000 files, 400 directories, 20000 modified paths; loose | 38.7 | 44.9 | 1.16× | ±0.84 ms |
| `tree-metadata-loose-cli-no-renames` | 120000 files, 400 directories, 20000 modified paths; loose | 38.3 | 43.6 | 1.14× | ±0.57 ms |
| `tree-metadata-cli-2083-packed-api` | 120000 files, 400 directories, 20000 modified paths; packed | 34.4 | 50.5 | 1.47× | ±2.60 ms |
| `tree-metadata-cli-2083-packed-cli-name-only` | 120000 files, 400 directories, 20000 modified paths; packed | 36.3 | 46.6 | 1.28× | ±1.40 ms |
| `tree-metadata-cli-2083-packed-cli-name-only-dist` | 120000 files, 400 directories, 20000 modified paths; packed | 36.3 | 44.3 | 1.22× | ±1.92 ms |
| `tree-metadata-cli-2083-packed-cli-name-status` | 120000 files, 400 directories, 20000 modified paths; packed | 34.7 | 47.0 | 1.35× | ±1.56 ms |
| `tree-metadata-cli-2083-packed-cli-name-status-dist` | 120000 files, 400 directories, 20000 modified paths; packed | 34.7 | 45.1 | 1.30× | ±1.26 ms |
| `tree-metadata-cli-2083-packed-cli-name-status-json` | 120000 files, 400 directories, 20000 modified paths; packed | 34.7 | 50.9 | 1.46× | ±1.51 ms |
| `tree-metadata-cli-2083-packed-cli-name-status-json-dist` | 120000 files, 400 directories, 20000 modified paths; packed | 34.7 | 46.7 | 1.34× | ±1.27 ms |
| `tree-metadata-cli-2083-packed-cli-name-status-no-renames` | 120000 files, 400 directories, 20000 modified paths; packed | 34.7 | 46.3 | 1.34× | ±3.10 ms |
| `tree-metadata-cli-2083-packed-cli-name-status-no-renames-dist` | 120000 files, 400 directories, 20000 modified paths; packed | 34.7 | 44.8 | 1.29× | ±1.40 ms |
| `tree-metadata-packed-api` | 120000 files, 400 directories, 20000 modified paths; packed | 25.7 | 51.0 | 1.98× | ±0.43 ms |
| `tree-metadata-packed-cli` | 120000 files, 400 directories, 20000 modified paths; packed | 26.5 | 39.4 | 1.49× | ±0.58 ms |
| `tree-metadata-packed-cli-no-renames` | 120000 files, 400 directories, 20000 modified paths; packed | 26.0 | 37.7 | 1.45× | ±0.26 ms |

### tree-metadata-criterion

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `tree-metadata-criterion-distributed-cold` | 120000 files, 400 directories, 20000 modifications; distributed subtree OIDs | — | 19.8 | — | ±0.00 ms |
| `tree-metadata-criterion-distributed-warm` | 120000 files, 400 directories, 20000 modifications; distributed subtree OIDs | — | 24.0 | — | ±0.00 ms |
| `tree-metadata-criterion-repeated-cold` | 120000 files, 400 directories, 20000 modifications; repeated subtree OIDs | — | 3.49 | — | ±0.00 ms |
| `tree-metadata-criterion-repeated-warm` | 120000 files, 400 directories, 20000 modifications; repeated subtree OIDs | — | 3.93 | — | ±0.00 ms |

### worktree

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `worktree-add-git-git` | git-git | 9.04 | 25.0 | **2.77×** | ±0.61 ms |
| `worktree-commit-git-git` | git-git | 17.6 | 38.4 | **2.18×** | ±1.00 ms |
| `worktree-diff-git-git` | git-git | 9.16 | 24.7 | **2.69×** | ±0.92 ms |
| `worktree-status-clean-git-git` | git-git | 6.83 | 25.3 | **3.70×** | ±0.71 ms |
| `worktree-status-dirty-git-git` | git-git | 8.83 | 26.6 | **3.01×** | ±0.65 ms |

### worktree-symlinks

| Scenario | Fixture | Git mean (ms) | Grit mean (ms) | Grit / Git | Spread |
| --- | --- | ---: | ---: | ---: | --- |
| `symlink-scan-regular-diff-clean` | 50,000 tracked regular files in 100 directories | 2,608 | 197 | 0.08× | ±1.06 ms |
| `symlink-scan-regular-diff-edited` | 50,000 tracked regular files in 100 directories | 2,605 | 197 | 0.08× | ±1.02 ms |
| `symlink-scan-regular-preview-clean` | 50,000 tracked regular files in 100 directories | 917 | 234 | 0.26× | ±0.86 ms |
| `symlink-scan-regular-preview-edited` | 50,000 tracked regular files in 100 directories | 915 | 366 | 0.40× | ±2.56 ms |
| `symlink-scan-mixed-diff-clean` | 50,000 tracked regular files in 100 directories + 64 directory symlinks | 2,632 | 199 | 0.08× | ±1.31 ms |
| `symlink-scan-mixed-diff-edited` | 50,000 tracked regular files in 100 directories + 64 directory symlinks | 2,611 | 199 | 0.08× | ±1.11 ms |
| `symlink-scan-mixed-preview-clean` | 50,000 tracked regular files in 100 directories + 64 directory symlinks | 920 | 241 | 0.26× | ±7.07 ms |
| `symlink-scan-mixed-preview-edited` | 50,000 tracked regular files in 100 directories + 64 directory symlinks | 920 | 371 | 0.40× | ±3.52 ms |


To refresh this page after updating a baseline JSON file, regenerate the static site:

```bash
make docs
```

The docs `--check` step fails if baseline numbers change without re-running `make docs`.

For other library surfaces under test, see the [grit-lib API on docs.rs](https://docs.rs/grit-lib).

## Filtered revision counts (2026-10-26)

An unlimited `grit revs --count` with author, committer or message filters now counts during traversal. It reads each visited commit once into a reusable buffer, uses borrowed commit headers and keeps delta bases cached, rather than retaining per-commit maps and sorting a list. Excluded ancestry, shallow boundaries, grafts and existing filter semantics still apply.

Factory VM: AMD EPYC 9V45, 4 vCPUs, Rust 1.99.0, `release` profile, system Git 2.55.0. Baseline `dbdf37d29719` and changed `c17f3c85ffb8` ran in the same sitting. The daily harness supplies the 50,000-commit / 20,000-file deep-history fixture and the upstream Git v2.47.0 fixture. After one warmup, ten interleaved runs per tool alternate Git, baseline grit and changed grit, reversing tool order on every other run; builds and tests were idle. Wall time is the median measured with Python `perf_counter` around `/usr/bin/time -v`; RSS is the maximum process peak across those runs. Raw wall/RSS samples are in `grit-utils/baselines/filtered-revision-count.json`.

| Scenario | Git ms | Grit before ms | Grit after ms | Grit peak RSS before KiB | Grit peak RSS after KiB |
| --- | ---: | ---: | ---: | ---: | ---: |
| `revs/count-author@deep-history` | 91.88 | 283.90 | 121.02 | 61,904 | 25,964 |
| `revs/author-n50@deep-history` | 2.17 | 3.63 | 3.64 | 17,332 | 17,408 |
| `log/author-n50@deep-history` | 2.10 | 3.56 | 3.58 | 17,724 | 17,792 |
| `revs/count-author@deep-history-cg` | 89.49 | 169.96 | 124.09 | 56,076 | 26,112 |
| `revs/author-n50@deep-history-cg` | 2.41 | 3.75 | 3.75 | 19,812 | 19,836 |
| `log/author-n50@deep-history-cg` | 2.35 | 3.64 | 3.65 | 19,876 | 20,096 |
| `revs/left-right-count@deep-history` | 2.75 | 5.40 | 5.46 | 15,784 | 15,740 |
| `revs/count@git-git` | 375.99 | 527.63 | 534.70 | 127,520 | 127,332 |
| `revs/count-author@git-git` | 479.37 | 917.16 | 385.81 | 129,796 | 59,256 |
| `revs/count-grep@git-git` | 486.45 | 885.70 | 366.93 | 123,720 | 59,244 |

The target drops from 3.09× to 1.32× Git time, with 57% less time and 58% less peak RSS. The other filtered counts improve by 27–59%; unchanged command paths differ by at most 1.4% in this run. Heaptrack records 1,247,228 → 266,825 allocations and 38.95 → 5.88 MB peak heap; callgrind records 4.38 → 2.14 billion instructions. `perf record -g` was blocked by the VM kernel setting, so callgrind and `strace -c` supplied CPU/syscall evidence.

`revs/count-path@git-git` could not run: system Git rejects the pinned `7e34d81` revision in the harness’s v2.47.0 fixture. The fixture and scenario definition were left unchanged.

### Commit-graph range counts follow-up

Filtered counts now use commit-graph parent links while marking excluded ancestry, and read object bodies for the included walk or for excluded commits missing from the graph. The reviewed implementation read all 49,500 excluded commit bodies for `HEAD~500..HEAD`; the new object-read regression test covers graph-enabled and disabled walks, merges and partial graph coverage. Shallow and grafted ranges are checked against system Git.

A second factory VM (Intel Xeon Platinum 8370C, 4 vCPUs, Rust 1.99.0, `release`, Git 2.55.0) compares origin/main `f0e7cddd530a`, reviewed branch `ca0542807699` and corrected code `0c235110d293` in the same sitting. The fixture and ten-run interleaved wall/RSS methodology match the earlier measurement; builds and tests were idle. Raw samples, command arguments and all three grit series are in `grit-utils/baselines/filtered-revision-count-range.json`. RSS below is the maximum process peak in KiB, from `/usr/bin/time -v`.

| Scenario | Git ms | Main ms | Reviewed ms | Fixed ms | Main RSS | Reviewed RSS | Fixed RSS |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| `log/author-n50@deep-history` | 2.59 | 5.05 | 5.25 | 4.77 | 15,968 | 16,100 | 17,036 |
| `revs/count-author@deep-history` | 151.22 | 533.64 | 207.67 | 204.06 | 60,872 | 24,968 | 25,880 |
| `revs/count-author-range@deep-history` | 4.24 | 325.87 | 201.85 | 196.57 | 54,132 | 25,240 | 26,068 |
| `revs/author-n50@deep-history` | 2.54 | 4.85 | 5.05 | 4.62 | 15,788 | 15,904 | 16,608 |
| `revs/count-grep-range@deep-history` | 4.13 | 326.81 | 201.35 | 195.83 | 53,912 | 25,040 | 25,800 |
| `revs/count@deep-history` | 112.74 | 6.96 | 6.61 | 6.06 | 15,200 | 15,232 | 16,424 |
| `log/author-n50@deep-history-cg` | 3.19 | 5.79 | 5.81 | 5.13 | 18,672 | 18,588 | 19,328 |
| `revs/count-author@deep-history-cg` | 159.65 | 312.30 | 208.71 | 205.47 | 55,072 | 25,168 | 26,004 |
| `revs/count-author-range@deep-history-cg` | 4.64 | 78.67 | 200.61 | 28.62 | 40,388 | 27,944 | 20,776 |
| `revs/author-n50@deep-history-cg` | 2.86 | 5.24 | 5.24 | 4.61 | 18,612 | 18,516 | 19,268 |
| `revs/count-grep-range@deep-history-cg` | 4.62 | 79.80 | 201.89 | 28.68 | 40,364 | 27,896 | 20,584 |
| `revs/count@deep-history-cg` | 23.45 | 6.77 | 6.89 | 6.75 | 15,444 | 15,428 | 16,372 |
| `revs/left-right-count@deep-history` | 3.74 | 8.62 | 8.71 | 8.07 | 14,132 | 14,464 | 15,408 |

The graph-backed author range count improves by 86% against the reviewed branch and 64% against main. Counts agree with Git, including 500 for the review reproduction. Full-history author counts retain the original optimization; graph loading is skipped when there are no exclusions. `perf record -g` remains unavailable (`perf_event_paranoid=4`); `strace -c` records about 1.4 ms in syscalls before the fix, while the deterministic graph-range test reduces object reads from 34 to 2.

This build records about 0.7–1.3 MiB more peak RSS on several unchanged paths than the reviewed build. A separate five-run `/proc/<pid>/smaps_rollup` sample of full-history author counting shows essentially unchanged peak anonymous memory (6,748 → 6,740 KiB), with the extra residency in file-backed mappings. Those samples are included in the baseline JSON; the RSS increases remain visible in the table rather than being treated as heap growth or omitted. No measured scenario slowed relative to main or the reviewed branch.

## Ranged blame pack residency

Issue #2079 measures `grit blame --rev 7e34d8174d -L 1,3 --json git-gui/git-gui.sh` on an unmodified clone of `git/git` (one 326 MB pack). The Git control is the system binary running `git blame --line-porcelain -L 1,3 7e34d8174d -- git-gui/git-gui.sh`.

Ranged blame now reads individual pack entries through positioned IO, including commit prefixes and delta bases, and keeps a private 16 MiB delta-base cache per object directory. Eight reusable 16 KiB compressed windows per reader coalesce nearby reads; larger entries bypass those windows. Each operation-local catalog retains at most 32 readers per object directory, evicting inactive readers and their descriptor/windows together. Thread-local caches use weak references for positioned readers, so fast lookups do not retain open files after eviction. Active reads retain their original file identity until completion. This covers ordinary ODB fallbacks, alternate stores and multiple packs as well as the session reader. Full-file blame and other ODB users retain their existing read/cache policy. Visited metadata, indexes and reconstructed objects still consume memory; the change bounds retained pack pages and delta bases rather than imposing a fixed total RSS limit.

The deterministic Linux regression uses 1,800 commits, 1,000 changing root-tree entries, a rename and a merge. It varies unrelated blob payloads from 8 KiB to 64 KiB per commit without changing the three requested lines. Before the fix, peak RSS increased from 81,488 to 128,816 KiB; positioned reads kept it near 57–58 MiB. `grit-bench history` includes the larger fixture as `history-blame-packed-range`. Tests create all objects locally with system Git and also check blame compatibility with and without a commit graph, single/multiple packs, and alternates.

Measured on a 4-core AMD EPYC 9V74 Linux VM using release builds, with one warmup per tool followed by seven rotated interleaved triplets (three for full-file blame). Builds and tests were paused during timing. Wall time includes fresh-process launch around `/usr/bin/time -v`; RSS is the median of each process's maximum. The frozen main build is `80dc322f7c2fafae948c30b9e6de159f884504e0`. Raw samples and commands are in `grit-utils/baselines/blame-range-memory-2079.json`.

| Operation | Main ms / KiB | Fixed ms / KiB | System Git ms / KiB |
| --- | ---: | ---: | ---: |
| Three-line blame | 449.52 / 331,872 | 434.26 / 135,744 | 171.61 / 83,468 |
| Full-file blame | 1248.65 / 389,408 | 1243.52 / 390,644 | 221.26 / 87,680 |
| Read commit | 3.46 / 13,408 | 3.22 / 13,576 | 2.30 / 5,308 |
| Read root tree | 9.17 / 28,720 | 8.90 / 28,696 | 2.59 / 5,392 |
| Read blob | 3.93 / 14,136 | 3.71 / 14,148 | 2.76 / 5,604 |

The ranged request uses **1.63× Git's RSS**, down **59.1%** from main, and is **3.4% faster** than main. Full-file blame and the object-read controls show no timing regression. The object-read controls have RSS differences below 0.2 MiB; raw samples retain these differences.

### Pack descriptor bound follow-up

The catalog now retains at most 32 pack readers per object directory, with weak thread-local references. Eviction closes inactive readers and releases their compressed windows; active readers continue using the same open file. A deterministic Rust regression puts historical commits, trees and changed target blobs in 300 Git-indexed packs (each over 8 KiB), then runs ranged blame with a 256-descriptor limit. The reviewed patch (`49047ae5`) fails with descriptor exhaustion; main, the fix and system Git succeed. Full-file blame remains a control, and the three attributed commits, original lines/paths and content match system Git.

Measured again on 2026-10-27, on an AMD EPYC 9V45 / 4-vCPU Ubuntu 24.04.5 VM, Rust 1.99.0 release builds and stock Git 2.55.0. Main is frozen at `e98de8a1`. One warmup precedes seven interleaved runs with rotating tool order, with builds and tests idle. Wall time wraps fresh subprocesses with a monotonic timer and RSS uses `/usr/bin/time -v`. Cells show median ms / median peak KiB. The 300-pack timing control uses the normal descriptor limit so the reviewed binary can complete. Raw samples are in the baseline's `descriptor_followup` field.

| Operation | Main | Reviewed patch | Fixed | System Git |
| --- | ---: | ---: | ---: | ---: |
| Three-line blame | 356.74 / 349,448 | 336.04 / 132,732 | 337.44 / 134,676 | 138.23 / 85,212 |
| Full-file blame | 805.06 / 404,524 | 808.72 / 406,124 | 802.26 / 405,580 | 174.61 / 89,288 |
| Read commit | 2.87 / 12,756 | 2.91 / 12,772 | 2.66 / 12,836 | 1.89 / 5,284 |
| Read root tree | 7.76 / 28,288 | 7.61 / 28,176 | 7.33 / 28,188 | 2.14 / 5,452 |
| Read blob | 3.36 / 13,164 | 3.36 / 13,524 | 3.08 / 13,568 | 2.41 / 5,476 |
| Three-line blame, 300 packs | 28.44 / 20,152 | 35.89 / 24,276 | 33.63 / 16,728 | 11.39 / 11,652 |

The exact git/git range peaks at **1.58× Git RSS**, **61.5% below main**, and runs **5.4% faster than main**. Its time is within 0.5% of the reviewed patch; full-file blame and representative ODB reads remain flat or improve. On 300 packs, reader eviction reduces RSS from 24,276 to 16,728 KiB and improves time by 6.3% against the reviewed patch.

## Parallel worktree symlinks (2026-10-27)

Parallel worktree scans now compare physical symlinks through the same link-target reader and file-type classifier as serial scans. Regular files retain parallel hashing and streaming. Commit previews also use symlink metadata when loading a new link blob into their temporary object overlay.

Factory VM: AMD EPYC 7763, 4 vCPUs, Rust 1.99.0, `release` profile, stock Git 2.55.0. The baseline is origin/main `80dc322f7`; fixed results are from this change. Each fixture has 50,000 tracked regular files in 100 directories; the mixed fixture adds 64 directory symlinks. The edited case changes one regular file and, for the mixed fixture, retargets one link.

One warmup precedes seven interleaved runs per tool, reversing tool order on alternating runs. Wall time is the median measured with Python `perf_counter` around `/usr/bin/time -v`; RSS is the maximum process peak across runs. Builds and tests were idle. A saved Git-written index with unset stat data is restored before every command to force content verification; these are **not trusted-stat clean-tree timings**. Commands are `git diff --no-ext-diff` versus `grit --threads 4 diff --json`, and `git commit -a --dry-run -m bench` versus `grit --threads 4 commit --dry-run -m bench --json`.

| Fixture / operation | Git ms | Main ms | Fixed ms | Git RSS KiB | Main RSS KiB | Fixed RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| regular diff, clean | 2623.28 | 198.22 | 196.33 | 26760 | 67188 | 67272 |
| regular preview, clean | 916.19 | 235.49 | 234.32 | 33420 | 73504 | 73804 |
| regular diff, edited | 2603.10 | 196.81 | 197.03 | 26768 | 67212 | 67216 |
| regular preview, edited | 913.63 | 366.93 | 365.37 | 33356 | 73816 | 73524 |
| mixed diff, clean | 2636.81 | 197.21 (error) | 198.56 | 26768 | 58680 | 67472 |
| mixed preview, clean | 918.62 | 214.50 (error) | 238.01 | 33424 | 64552 | 74624 |
| mixed diff, edited | 2612.90 | 196.00 (error) | 198.64 | 26772 | 58436 | 68336 |
| mixed preview, edited | 918.81 | 215.26 (error) | 372.02 | 33528 | 64988 | 74860 |

Regular-file timings stay within 1% of main and peak RSS within 0.5%. Mixed-fixture main results are directory-read errors; their time and RSS measure early failure and cannot be compared as successful scans. Fixed mixed diffs succeed; clean previews return `nothing_to_commit`, and edited previews succeed. Raw samples, exit codes, command arguments and fixture details are in `grit-utils/baselines/diff-parallel-symlinks-2076.json`.

The existing 50,000-file performance suite also passes all four checks with the fixed release CLI: warm clean diff is 148.49 ms versus Git 95.56 ms (1.55×), 100 edits are 150.88 versus 156.11 ms (0.97×), and mtime-only touches are 146.03 versus 94.83 ms (1.54×). The deletion parity check passes. These existing fixtures include deep paths and retain their original 3× timing limit.

## Staging with case-insensitive path settings (issue #1682, 2026-10-27)

`git init` on macOS writes `core.ignorecase=true` and `core.precomposeunicode=true`, and Windows sets `core.ignorecase=true`. On case-folding volumes, `grit add` and commit auto-staging keep the index's path spellings instead of rewriting them from directory listings (issue [#2247](https://github.com/gitbutlerapp/grit/issues/2247)); explicit case renames use [`grit mv`](https://grit-scm.com/docs/commands/mv.md/index.md). On case-sensitive volumes with `core.ignorecase=true`, staging still refreshes case-only renames from listings when the indexed spelling no longer opens. The listing cache answers a path that exists exactly as indexed from its parent's listing without allocating, reads every parent directory up front in parallel, and stores listing names in one buffer. Case-collision groups keep only paths that actually collide. ASCII paths skip the precompose lookup, which stat'ed every path before returning it unchanged. Directory pruning for literal pathspecs no longer allocates per directory and pathspec.

Factory VM: Intel Xeon, 4 vCPUs, Rust 1.99.0, `release` profile, stock Git 2.43.0. The fixture is the issue's tree: `d{0..6}/e{0..6}/f{0..6}/g{0..6}` with 20 files per leaf, 48,020 tracked files, 98 of them modified. The baseline is origin/main `1dea9fde5`, which already caches directory listings per pass ([#2102](https://github.com/gitbutlerapp/grit/issues/2102)). Every command starts from an index reset to the base commit; commit rows also run `git add -A` first and compare `grit commit -m x` with `git commit -m x`. One warmup per tool precedes seven interleaved runs, reversing tool order every other run. Wall time is the median measured with Python `perf_counter` around `/usr/bin/time -v`, and RSS is the maximum peak across runs. Builds and tests were idle. Raw samples are in `grit-utils/baselines/add-ignorecase-1682.json`.

| Config / operation | Git ms | Main ms | Fixed ms | Git RSS KiB | Main RSS KiB | Fixed RSS KiB |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| ignorecase, add 1 path | 28.0 | 143.9 | 53.5 | 16,384 | 58,196 | 38,232 |
| ignorecase, add 10 paths | 31.2 | 142.1 | 50.6 | 16,380 | 57,824 | 36,512 |
| ignorecase, add 98 paths | 72.6 | 166.8 | 60.8 | 16,404 | 60,648 | 40,744 |
| ignorecase, `add .` vs `git add -A` | 56.4 | 290.0 | 107.8 | 15,920 | 61,688 | 36,956 |
| ignorecase, `add -u` | 39.2 | 252.3 | 66.8 | 15,920 | 61,160 | 35,736 |
| ignorecase, `add --dry-run .` | 59.6 | 234.8 | 87.6 | 15,920 | 63,176 | 30,856 |
| ignorecase, commit | 54.2 | 323.7 | 138.6 | 17,236 | 61,492 | 37,544 |
| ignorecase, status (control) | 39.4 | 53.4 | 46.4 | 16,672 | 32,688 | 32,844 |
| default, add 1 path | 25.5 | 30.0 | 30.5 | 15,920 | 27,520 | 27,308 |
| default, add 10 paths | 28.0 | 29.8 | 28.7 | 15,928 | 27,352 | 27,512 |
| default, add 98 paths | 70.4 | 51.0 | 36.5 | 15,888 | 30,120 | 29,708 |
| default, `add .` vs `git add -A` | 50.8 | 81.5 | 83.0 | 15,920 | 32,852 | 32,488 |
| default, `add -u` | 34.0 | 41.5 | 41.5 | 15,876 | 29,620 | 29,276 |
| default, `add --dry-run .` | 53.9 | 59.7 | 59.1 | 15,916 | 26,312 | 26,252 |
| default, commit | 49.8 | 117.8 | 110.6 | 17,256 | 34,744 | 35,164 |
| default, status (control) | 38.2 | 41.9 | 43.2 | 16,672 | 32,148 | 32,364 |

With the macOS settings, staging and commit are 2.3–3.8× faster than main and use 34–51% less peak memory. Extra paths now cost little: 98 paths take 60.8 ms against 53.5 ms for one, faster than Git's 72.6 ms, though a single path is still 1.9× Git. `add .` and `add -u` remain 1.7–1.9× Git. Default-config rows stay within run-to-run noise, apart from 98 explicit paths (28% faster, from the pathspec pruning). `grit commit` still stats every tracked file and walks the tree for untracked files because it stages everything by default; `git commit` writes the existing index only. Before this change on the earlier main `ad7f51066`, the same `add .` took 1,102 ms and `add -u` 1,026 ms.

On a 48k-file tree, `add_many_pathspec_perf_git_compat` checks that explicit paths, `add .` and `add -u` under these settings stay within 5× of system Git. Daily scenarios `add/all-ignorecase@synth-100k` and `add/update-ignorecase@synth-100k` follow the same path on the standard 100k fixture.

## Short revision listings (2026-10-28)

`grit revs --author=. -n 50 HEAD` on the deep-history fixture spent most of its time on fixed costs, not on walking 50 commits. Peak heap is about 380 KB; the rest of the gap to Git is the binary's own pages: relocation and code page faults, clap setup, repository discovery. The changes remove work from that fixed part:

- Human and Markdown output from `grit revs` print each commit's id, subject and side marks. The command no longer loads ref decorations (every ref, with tags peeled) or full JSON rows (identities, RFC 3339 and relative dates, change ids) for them; `--json` output is unchanged.
- Opening a repository no longer copies the whole process environment for filter subprocesses. The copy is made the first time a clean or smudge filter runs.
- `ObjectId::to_hex` and `Display` write hex digits from a lookup table instead of the `hex` crate's per-char iterator; `Display` no longer allocates.

For the target, callgrind records 7.04 → 5.92 million instructions, heaptrack 6,408 → 4,996 allocations, and `perf stat` 581 → 556 page faults.

Factory VM: Intel Xeon, 4 vCPUs, Rust 1.99.0, `release` profile (grit-cli and grit-utils built together), system Git 2.43.0. Baseline `74d8bf801` and changed `1d7ea0491` ran in the same sitting. After three warmups, the three tools ran interleaved with rotating order: 300 rounds, or 15 for the long count rows. Wall time is the median measured from spawn to exit. Peak RSS is the median of seven `/usr/bin/time` runs, in KiB. `revs-json` is the target with `--json`. Raw samples are in `grit-utils/baselines/revs-startup-4992.json`.

| Scenario | Git ms | Grit before ms | Grit after ms | Git RSS | Grit RSS before | Grit RSS after |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| `revs/author-n50@deep-history` | 1.48 | 3.08 | 2.88 | 6,656 | 17,444 | 17,008 |
| `revs/author-n50@deep-history-cg` | 1.76 | 3.16 | 2.95 | 9,944 | 20,060 | 19,492 |
| `revs/count-author@deep-history` | 113.14 | 123.80 | 124.85 | 19,084 | 26,680 | 26,532 |
| `revs/count-author@deep-history-cg` | 123.06 | 123.38 | 123.76 | 22,276 | 26,824 | 26,316 |
| `revs/left-right-count@deep-history` | 2.17 | 3.99 | 3.96 | 6,836 | 16,004 | 15,920 |
| `revs/count@git-git` | 516.95 | 657.44 | 662.23 | 59,312 | 125,648 | 125,532 |
| `log/author-n50@deep-history` | 1.50 | 3.08 | 3.06 | 6,704 | 17,592 | 17,436 |
| `log/author-n50@deep-history-cg` | 1.78 | 3.20 | 3.17 | 10,004 | 19,960 | 19,760 |
| `log/json-100@git-git` | 2.78 | 4.37 | 4.35 | 12,572 | 21,676 | 21,820 |
| `log/oneline-100@git-git` | 2.83 | 4.38 | 4.34 | 12,664 | 22,884 | 22,868 |
| `show/head@git-git` | 1.70 | 2.70 | 2.71 | 6,796 | 16,216 | 16,536 |
| `revs-json/author-n50@deep-history` | 1.45 | 3.03 | 3.03 | 6,704 | 17,612 | 17,676 |

The target is 6.4% faster with 436 KiB less peak RSS; with a commit-graph it is 6.6% faster with 568 KiB less. Other rows stay within ±1% wall time, the run-to-run noise of these measurements. Peak RSS of the release binary moves by up to ±400 KiB between builds of unrelated code, because code placement decides which pages a short command touches. The `show/head` RSS increase is in that range.

A separate gap turned up while measuring: `revs --count --since=2.weeks HEAD` without `-n`, on this fixture where every commit is older than the cutoff, takes 235 ms against Git's 1.7 ms. grit walks all 50,000 commits, while Git stops after a few. This change does not affect that path.

What remains is mostly outside the walk. Applying the binary's 38,826 relative relocations at startup copies every `.data.rel.ro` page (about 100 page faults). In a separate comparison of two builds of the same code, a non-PIE build ran the target in 2.35 ms instead of 2.63 ms, with 1.3 MiB less RSS, at the cost of ASLR for the executable. Commit reads cost about 27,000 instructions each, of which about 10,000 are inflate (as in Git); most of the rest is pack-cache bookkeeping keyed by pack path. Translating the `.` in `--author=.` produces a long byte-class alternation, and parsing it accounts for most of the 300,000 instructions spent building the filter.

## Branch listing and clap's global flags (2026-10-29)

`grit branch` on the many-refs fixture lists one local branch, so nearly all of its time is fixed cost. On the `dist` binary, callgrind counts 2.16 million instructions against Git's 0.96 million, and `grit --version` alone takes 1.30 million. The largest single item after the dynamic loader was argument parsing. Before every parse, clap copies each global flag of the root command (`--json`, `--markdown`, `--filter`, `--threads`) into every subcommand, so 55 subcommands received 220 argument copies that were built and dropped again, although only the invoked subcommand can use them.

grit now parses with a command tree in which only the subcommand named in `argv` receives those flags, in the same order clap would add them, and the root stops marking them global. Without a known subcommand name, as with `grit`, `grit --help` or a misspelled command, the stock tree is used, so suggestions still list every command. A unit test parses every subcommand and alias with the flags before and after it, nested subcommands, conflicting flags, missing values, unknown flags and `--help`, and requires the same parsed values and the same rendered errors as the stock tree.

For the target (`grit branch`, `dist` profile, isolated config), callgrind records 2.16 → 1.76 million instructions (−18.6%), heaptrack 2,346 → 1,622 allocations and 339 → 177 KB peak heap, and `perf stat` 400 → 362 page faults. `grit branch --json` drops from 2.35 to 1.95 million instructions.

Factory VM: Intel Xeon, 4 vCPUs, Rust 1.99.0, `dist` profile, system Git 2.43.0. Baseline `71dda1c22` and the change ran in the same sitting with `GIT_CONFIG_NOSYSTEM=1` and a near-empty global config. After five warmups, the three tools ran interleaved with rotating order for 400 rounds. Mutating rows were reset with Git, untimed, before each sample. Wall time is the median from spawn to exit. Peak RSS is the median of 21 `/usr/bin/time` runs, in KiB. Raw samples are in `grit-utils/baselines/clap-globals-5562.json`, and two `grit-bench daily` runs with the tools in both orders are in `daily-clap-globals-5562-before.json` and `daily-clap-globals-5562-after.json`.

| Scenario | Git ms | Grit before ms | Grit after ms | Git RSS | Grit RSS before | Grit RSS after |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| `branch/list@many-refs` | 1.13 | 1.72 | 1.71 | 4,912 | 11,068 | 11,008 |
| `branch/list-json@many-refs` | 1.30 | 1.87 | 1.86 | 5,680 | 11,892 | 11,848 |
| `branch/list-markdown@many-refs` | 1.31 | 1.88 | 1.88 | 5,612 | 11,668 | 11,672 |
| `branch/list@many-refs-loose` | 1.13 | 1.72 | 1.71 | 4,912 | 11,124 | 10,984 |
| `branch/merged@many-refs` | 1.25 | 1.91 | 1.89 | 5,532 | 12,008 | 11,920 |
| `branch/create@many-refs` | 1.39 | 2.21 | 2.19 | 5,684 | 11,932 | 11,844 |
| `branch/delete@many-refs` | 4.06 | 4.44 | 4.22 | 5,700 | 11,780 | 11,684 |
| `tag/list@many-refs` | 2.72 | 3.79 | 3.79 | 5,820 | 11,576 | 11,576 |
| `config/get@small` | 0.98 | 1.60 | 1.57 | 4,284 | 9,884 | 9,752 |
| `status/clean@small` | 1.45 | 2.28 | 2.24 | 5,168 | 12,716 | 12,844 |
| `log/author-n50@deep-history` | 1.47 | 2.55 | 2.54 | 6,760 | 14,592 | 14,620 |
| `grit --version` | 0.87 | 1.12 | 1.16 | 3,944 | 5,900 | 5,956 |

Wall time improves by about 1% on most rows (`branch -D` by 5%). The instruction savings are larger than the time savings because the work removed was cheap allocation and copying, while page faults dominate the remaining time. `grit --version` runs no subcommand, so the change leaves it untouched: it runs 1,300,127 instructions before and 1,298,556 after, and its 3 extra page faults come from code placement in the new build. RSS rows move by up to ±150 KiB for the same reason.

The rest of the gap is in how the binary is loaded and laid out, not in branch listing. ld.so spends 690,000 instructions and about 100 copy-on-write page faults applying 39,608 relative relocations, and resolving symbols. `grit branch` then faults in code from about 90 separate 64 KiB windows spread across 25 MB of `.text`, which accounts for most of the 6 MB of resident code (Git: 1.5 MB). As a prototype, the same build was relinked with LLD's `--symbol-ordering-file` listing the 871 functions that callgrind saw in nine common commands. That cut `grit branch` peak RSS from 10.7 to 7.5 MB, its page faults from 358 to 304 and its task clock from 1.76 to 1.44 ms; `status` and `log -n 5` each lost about 2.6 MB of RSS. Symbol names include crate hashes that change with every version and toolchain, so an ordering file has to be regenerated as part of each build. That step is not part of this change.
## Config loading and revalidation (2026-10-29)

`grit branch --json` on the many-refs fixture lists one local branch, so it measures fixed costs. After process startup and argument parsing, the largest library cost was configuration. Parsing the system, global and repository files took about 3,800 instructions per entry: every line was copied, the inline-comment stripper built a new string, each key lowercased its section again through `format!`, and the quote and continuation scanners decoded characters. Every `Repository::config()` call then revalidated the cached cascade by rebuilding the environment fingerprint, the load options and the list of cascade paths before statting them, about 20,000 instructions per call. Listing branches makes four such calls, and `status` makes many more.

The parser now borrows each line unless a continuation joins it with the next, returns the comment-stripped value as a slice of the input, keeps the canonical section prefix per header so building a key is one allocation, and scans for the ASCII delimiters byte by byte. A unit test checks the byte scanners against the previous character scanners on every string of up to five tokens drawn from quotes, backslashes, comment characters, whitespace and multi-byte text. The config session remembers the environment snapshot it was last validated against. While the handle still uses that snapshot and git dir, the fingerprint, cache key and paths cannot have changed, so a check only re-stats the stamped files. A different snapshot gets the full check.

The Criterion bench `config_parse` (in `grit-lib/benches`) measures parsing a 40-entry user config at 11.55 → 7.26 µs, `canonical_key` at 244 → 119 ns for two keys, and a repeated `Repository::config()` at 3.02 → 1.60 µs, the rest being the `stat` calls.

Factory VM: Intel Xeon, 4 vCPUs, Rust 1.99.0, `dist` profile, system Git 2.43.0. The baseline is `e84d7c9df`. Each row ran system Git, baseline grit and changed grit interleaved, with alternating order, for 600 rounds after ten warmups. Wall time is the median from spawn to exit. Peak RSS is the median of 21 `/usr/bin/time` runs, in KiB. Instructions are callgrind totals for one run. "User config" is the VM's own `/etc/gitconfig` and a 1.1 KB `~/.gitconfig` (36 entries in all); "isolated" sets `GIT_CONFIG_NOSYSTEM=1` and an empty `GIT_CONFIG_GLOBAL`.

| Scenario | Config | Git ms | Grit before ms | Grit after ms | Git RSS | Grit RSS before | Grit RSS after | Instructions before → after |
| --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| `branch/list-json@many-refs` | user | 1.32 | 1.96 | 1.93 | 5,588 | 12,300 | 11,880 | 2.71M → 2.55M |
| `branch/list-json@many-refs` | isolated | 1.24 | 1.90 | 1.89 | 5,568 | 12,272 | 11,888 | 2.46M → 2.36M |
| `branch/list@many-refs` | user | 1.17 | 1.83 | 1.79 | 4,848 | 11,300 | 10,936 | 2.53M → 2.37M |
| `branch/list@many-refs-loose` | user | 1.14 | 1.82 | 1.81 | 4,880 | 11,256 | 10,784 | 2.54M → 2.38M |
| `branch/merged@many-refs` | user | 1.26 | 2.01 | 2.00 | 5,504 | 12,148 | 11,788 | 2.78M → 2.60M |
| `config/get@small` | user | 0.97 | 1.71 | 1.70 | 4,232 | 10,084 | 9,916 | 2.74M → 2.55M |
| `status/clean@small` | user | 1.40 | 2.32 | 2.20 | 5,172 | 12,928 | 12,608 | 5.52M → 3.46M |
| `status/clean@small` | isolated | 1.36 | 2.29 | 2.25 | 5,164 | 13,008 | 12,660 | 3.45M → 3.16M |

Branch listing runs 4–6% fewer instructions and here used about 380 KiB less memory, but its wall time moves by only 0–2%: most of the remaining gap to Git is loading the binary (relocations and code pages spread over `.text`), which this change does not touch. `status` gains the most, since it reads configuration many times per run. `grit-bench daily` runs of the branch, config, status and tag rows with each binary are in `daily-config-parse-5571-before.json` and `daily-config-parse-5571-after.json`; those runs are not interleaved, and their rows move by up to ±7% between runs in both directions.

After rebasing onto `3b7d852ca` (which includes the clap change above), callgrind counts with the user config are 2.21M → 2.06M instructions for `branch --json` (−7.0%), 2.02M → 1.87M for `branch` (−7.5%), 2.27M → 2.10M for `branch --merged main` (−7.7%), 2.25M → 2.06M for `config` (−8.4%) and 3.25M → 2.89M for `status` (−11.1%). Page faults are unchanged (381 for `branch --json`, 423 for `status` with a 360-line global config). At this size, wall time and peak RSS do not resolve the change: two `dist` builds of the same baseline source, one stripped and one with symbols, differ by up to 4% in wall time and 500 KiB in peak RSS in the same interleaved run, because shifting `.text` changes which 64 KiB windows the start-up path faults in. Changes to these rows below about 4% need instruction counts or a layout-independent measure.

## Setting and reading one config key (2026-10-31)

`grit config <key> <value>` used to open the whole repository to find `.git/config`. That meant the object database and the ref store (`packed-refs` and `refs/replace` probes), plus a config session that stamped and re-stamped every cascade file. It then fsynced the rewritten file before the rename. Reads loaded the cascade twice: once during repository discovery and once for the lookup.

Config commands now locate the repository with [`Repository::discover`](https://docs.rs/grit-lib/latest/grit_lib/repo/struct.Repository.html) and take paths from [`RepositoryLocation::from`](https://docs.rs/grit-lib/latest/grit_lib/repo/struct.RepositoryLocation.html) on the handle. That loads the merged config cascade once on the open repository (including `core.bare`, `core.worktree`, and `-c` / `GIT_CONFIG_*` overrides). The write itself no longer fsyncs. Git has no `core.fsync` component for config files, so it does not fsync config either, and the lock file and atomic rename are unchanged. A set now rewrites only the target file.

Factory VM: Intel Xeon, 4 vCPUs, Rust 1.99.0, `release` profile, system Git 2.43.0, ext4 scratch disk, baseline `3c14b849f`. These are `grit-bench daily --only config/set@small,config/get@small,config/get-origin@small` runs with 10 warmups and at least 100 timed runs. The before and after binaries ran back to back, twice each in alternating order. Medians are in milliseconds and peak RSS is in MiB, from the second pair (`daily-config-set-before.json`, `daily-config-set-after.json`); the first pair agrees within 0.12 ms.

| Scenario | Git ms | Grit before ms | Grit after ms | Git RSS | Grit RSS before | Grit RSS after |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| `config/set@small` | 1.32 | 3.09 | 2.08 | 6.3 | 13.2 | 11.9 |
| `config/get@small` | 1.16 | 2.04 | 1.87 | 6.1 | 13.3 | 11.9 |
| `config/get-origin@small` | 1.16 | 2.03 | 1.90 | 6.2 | 13.0 | 12.1 |

About 0.7 ms of the set improvement is the dropped fsync, and the rest is the skipped object database, ref store and config-session stamping. `grit config --file .git/config`, which needs no discovery, went from 2.3 to 1.6 ms in a separate hyperfine run. The remaining gap to Git is process startup: `grit --version` alone takes about 1.4 ms on this VM, against 0.9 ms for a complete `git config` set ([PIE startup relocation probe](#pie-startup-relocation-probe)).

## Primary store dispatch (task 7286, 2026-11-07)

An `Odb` used to keep its primary store in two fields: an optional custom store and a lazily opened files source. Every object read asked for the primary twice. It cloned one `Arc` to get the store as a `dyn WritableObjectStore`, cloned the files source again for the corrupt-loose-file check, and then read through the trait object. The primary is now one internal enum with a files arm and a custom arm. A read borrows the opened store, calls the files source directly (no `dyn` call, no reference counting) and reaches a custom store through its trait object as before. The public API is unchanged.

Factory VM: Intel Xeon, 4 vCPUs, Rust 1.99.0, system Git 2.43.0. The baseline is `97fd8dac8`. CLI rows used `dist` builds of both, with `GIT_CONFIG_NOSYSTEM=1` and an empty global config. After five warmups, system Git, baseline grit and changed grit ran interleaved, rotating the order each round, with no builds or tests running. Wall time is the median from spawn to exit. The change column is the median of per-round after/before ratios, with a bootstrap 95% interval. Raw samples, instruction counts and page faults are in `grit-utils/baselines/primary-store-7286.json`. Two `grit-bench daily` runs, one per binary, are in `daily-primary-store-7286-before.json` and `daily-primary-store-7286-after.json`. Those runs are not interleaved, so they move with the machine: Git itself differed by up to 11% between them.

| Scenario | Rounds | Git ms | Grit before ms | Grit after ms | Change |
| --- | ---: | ---: | ---: | ---: | --- |
| `log/oneline-100@git-git` | 600 | 3.22 | 3.90 | 3.90 | +0.2% (−0.5 to +0.5) |
| `show/head@git-git` | 600 | 2.22 | 2.70 | 2.70 | −0.5% (−1.1 to +0.6) |
| `status/dirty@synth-10k` | 400 | 7.38 | 7.17 | 7.23 | +1.4% (+0.7 to +2.1) |
| `diff/worktree@synth-10k` | 400 | 6.85 | 6.56 | 6.66 | +1.4% (+0.5 to +2.5) |
| `blame/file@git-git` | 60 | 292.8 | 253.3 | 251.6 | −0.7% (−1.3 to +0.3) |
| `revs/count@git-git` | 45 | 525.8 | 388.4 | 381.8 | −2.0% (−3.2 to −1.3) |
| `revs/count@synth-120k` | 45 | 385.3 | 265.8 | 256.0 | −3.4% (−4.1 to −2.8) |

`synth-120k` is a 120,000-commit linear history built with `git fast-import` and repacked without a commit-graph, so `grit revs --count HEAD` reads every commit through the primary store. It gains the most, at about 10 ms. `status` and `diff` do no more work than before. Callgrind counts the same instructions or fewer for every CLI row (`diff` 41.96M → 41.94M, `status` 47.94M → 47.93M, `log --limit 100` 11.844M both, `show` 3.636M both, `status` on a 100k-file worktree 341.18M → 340.94M). Minor page faults drop by two for each (`diff` 810 → 808, `status` 794 → 792). An earlier interleaved run of the same two rows with the first build of the change measured +0.7% for both. Moves of this size come from code placement in the fat-LTO binary, not from the change.

Per-operation instruction counts for `Odb` calls on a files primary were measured as the difference between 2,000 and 4,000 repeated calls over 64 blobs in a `release` build, under callgrind. "In context" means inside `Odb::with_pack_read_context`.

| Operation | Packed before | Packed after | Packed in context before | Packed in context after | Loose before | Loose after |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| `read` (session cache hit) | 493 | 493 | — | — | 494 | 494 |
| `read` (pack batch path) | — | — | 10,235 | 10,229 | — | — |
| `read_into` | 11,127 | 11,065 | 9,999 | 9,994 | 24,675 | 24,613 |
| `read_info` | 5,913 | 5,849 | 5,130 | 5,127 | 18,895 | 18,821 |
| `contains` | 6,039 | 6,024 | 5,669 | 5,654 | 2,063 | 2,048 |
| `exists_local` | 2,180 | 2,180 | 1,749 | 1,749 | 3,699 | 3,700 |

The first build of the change made the session-cache hit in `read` 11 instructions dearer (493 → 504). Callgrind traced this to `SessionLooseReads::get` no longer being inlined into the cache lookup. In `release` and `bench` builds (16 codegen units, no LTO), whether that small function is inlined depends on how the crate is split into codegen units, and adding a module had moved it. It now carries `#[inline]`, which brings the hit back to 493. The shipped `dist` build uses one codegen unit and was not affected. The same partitioning explains why benchmarks of functions this change does not touch moved by several percent: `odb_read/delta_apply` measured +6.9% while running exactly the same 515 instructions per call.

The Criterion suites `odb_read`, `odb_backend`, `odb_backend_pack`, `odb_read_parallel` and `hot_paths` ran six times each, alternating base and change. The table gives the median of the three runs per side.

| Benchmark | Before | After | Change |
| --- | ---: | ---: | ---: |
| `odb_backend/read/pack_100k` | 32.2 ns | 32.5 ns | +1.0% |
| `odb_backend/read_info/pack_100k` | 663 ns | 663 ns | −0.1% |
| `odb_backend/exists_hit/pack_100k` | 855 ns | 834 ns | −2.5% |
| `odb_backend/read/loose_10k` | 30.5 ns | 30.9 ns | +1.3% |
| `odb_backend/read_info/loose_10k` | 3.42 µs | 3.39 µs | −0.9% |
| `odb_backend/read_info/midx_8` | 522 ns | 505 ns | −3.2% |
| `odb_backend/read_info/alternate_only` | 5.33 µs | 5.32 µs | −0.2% |
| `odb_backend/read_info_packed` | 472 ns | 473 ns | +0.2% |
| `odb_read/packed_read/deep_delta_chain` | 43.0 µs | 42.9 µs | −0.4% |
| `odb_read/deep_repack_read_one` | 33.2 µs | 33.1 µs | −0.3% |
| `checkout_between_trees/H_0.01` | 41.7 ms | 40.1 ms | −3.7% |
| `checkout/switch_wide_100k/H_10pct_switch` | 428 ms | 426 ms | −0.6% |
| `diff/worktree_clean_10k/grit_lib` | 3.63 ms | 3.67 ms | +1.0% |
| `index_mutate_batch_10pct/H` | 6.87 ms | 6.78 ms | −1.3% |
| `staging_scan_5_modified/H` | 60.6 ms | 57.8 ms | −4.7% |

Some rows are left out of the table:

- The `write/*` rows grow the object database on every iteration and swung by up to ±70% between runs of the same binary.
- `odb_read_parallel` swung by ±20% between runs of the same binary.
- `checkout_between_trees/H_0.1` (+4.3%) and `status/clean_100k` (+3.2%) each had one disturbed run. An eight-run rerun of just those two gave +0.2% for the checkout. The status benchmark stats 100,000 files and stayed within its own spread: its four baseline runs alone ranged from 26.3 to 30.9 ms. Callgrind counts 341.18M instructions before and 340.94M after for `grit status` on that tree.

The `hot_paths` suite stops at its stash benchmark on this baseline too (`apply_stash` reports `WouldOverwriteLocalChanges`), so its last five benchmarks were not measured. A new test, `grit-lib/tests/odb_primary_read_allocations.rs` (`full-tests`), holds warm reads to the heap allocations both builds make: 1 per packed `read_into`, 2 per packed `read_info`, 10 per packed `contains`, none inside a pack read context, and 15, 14 and 3 for the loose equivalents.


## Text diffs through one options struct (2026-11-08)

[`unified_diff`](https://docs.rs/grit-lib/latest/grit_lib/diff/fn.unified_diff.html) and [`line_change_counts`](https://docs.rs/grit-lib/latest/grit_lib/diff/fn.line_change_counts.html) now take one [`UnifiedDiffOptions`](https://docs.rs/grit-lib/latest/grit_lib/diff/struct.UnifiedDiffOptions.html) and render every option through the engine `write_patch` uses. Before, each option had its own renderer: one per algorithm, a separate `-W` engine built on the `similar` crate, and an `--anchored` diff that split the file at exact-match lines.

The rows come from three Criterion benches in `grit-lib/benches`:

- `diff`:
  - `diff/blob_myers_10k_lines` uses the shared fixture's 10,000-line blob plus one appended line.
  - `diff/text_count` (line counts, the `--stat` path) and `diff/text_render` (full patches with `a/` / `b/` labels and function names) use a generated 10,000-line Rust file: 2,500 small functions, with an edit every 37, a re-indented body every 53 and an inserted function every 101.
  - The pathological rows use the shared fixture's 4,000-line `x`/`y` alternation.
- `userdiff_hunk_header` times one rust-driver funcname match.
- `merge_content` times a 20,000-line three-way content merge.

The baseline is main at `e1bbc8635`, running the same `diff` bench file with the calls routed through a local shim over the old public functions (`count_changes_with_algorithm`, `unified_diff_with_prefix_and_funcname_and_algorithm`, `anchored_unified_diff`, `line_diff_unified_hunk_body`). Both sides build the same inputs in the same order.

Factory VM: Intel Xeon, 4 vCPUs, Rust 1.99.0, `bench` profile. Criterion took 40 samples per row (100 for `userdiff_hunk_header`) in three rounds, alternating baseline and branch binaries. The table shows the median of the three round medians. One branch round ran during a slow spell, where every row was 7–20% slower, including tree diffs this change does not touch. The rows near 3% got five more alternating rounds with 60 samples, noted below.

| Benchmark | Before | After | Change |
| --- | ---: | ---: | ---: |
| `userdiff_hunk_header_rust_match_line` | 282–294 ns | 270–277 ns | −4.5% |
| `merge_trees_three_way/content_both_sides_20k` | 745.0 ms | 749.3 ms | +0.6% |
| `diff/blob_myers_10k_lines/myers` | 315 µs | 315 µs | +0.2% |
| `diff/blob_myers_10k_lines/histogram` | 407 µs | 411 µs | +0.9% |
| `diff/text_count/code_myers` | 455 µs | 466 µs | +2.5% (five extra rounds) |
| `diff/text_count/code_minimal` | 422 µs | 425 µs | +0.9% |
| `diff/text_count/code_patience` | 1.891 ms | 1.853 ms | −2.0% |
| `diff/text_count/code_histogram` | 1.511 ms | 1.502 ms | −0.6% |
| `diff/text_count/pathological_myers` | 186 µs | 189 µs | +1.7% |
| `diff/text_count/pathological_histogram` | 76.7 ms | 75.8 ms | −1.2% |
| `diff/text_render/myers` | 703 µs | 601 µs | −14.4% |
| `diff/text_render/minimal` | 673 µs | 574 µs | −14.8% |
| `diff/text_render/patience` | 2.138 ms | 2.009 ms | −6.1% |
| `diff/text_render/histogram` | 1.897 ms | 1.782 ms | −6.1% |
| `diff/text_render/u0` | 705 µs | 600 µs | −14.8% |
| `diff/text_render/inter_hunk_6` | 713 µs | 598 µs | −16.1% |
| `diff/text_render/funcname_rust` | 802 µs | 686 µs | −14.5% |
| `diff/text_render/function_context` (`-W`) | 2.483 ms | 719 µs | −71.0% |
| `diff/text_render/anchored` | 2.592 ms | 2.199 ms | −15.2% |
| `diff/text_render/ignore_space_change` (`-b`) | 1.728 ms | 1.799 ms | +4.1% (not like for like, see below) |
| `diff/tree_diff_wide/few` | 345 µs | 350 µs | +1.4% (five extra rounds) |
| `diff/tree_diff_siblings/distributed/warm` | 17.81 ms | 17.25 ms | −3.2% (five extra rounds) |

**Funcname matching.** On the first rebased build, `userdiff_hunk_header` ran about 4% slower than main in every alternating round, though `FuncnameMatcher::match_line` had not changed. Callgrind showed the cause: after the consolidation, the compiler no longer inlined `regex::Regex::captures_at` into `match_line`, which added 34 instructions per call (4,610 against 4,576). `match_line` also did two avoidable things. It ran a second search (`is_match` after `captures` had already found no match, which cannot succeed), and it built the matched text twice (a `String` from `trim_end_matches(char::is_whitespace)`, then a second one from `trim_end`). One search and one trimmed copy return the same text in 4,405 instructions per call, 3.7% fewer than main. Wall time is now 4–6% faster than main, and all 31 built-in driver cases still match `git diff`'s hunk headers byte for byte.

**Counting rows.** They move in both directions between rounds. Per call, callgrind measures the same or slightly less work than main: Myers counting 8.589M against 8.607–8.618M instructions, histogram 0.6% fewer, minimal 0.25% fewer. The wall-time spread follows branch mispredictions in the line interner's `foldhash` hashing, which `foldhash::fast::RandomState` seeds differently in every process. Two runs of the same main binary gave 35.0k and 37.9k mispredictions per Myers-count call, against 34.7k and 36.3k for the branch. Plain patience keeps its original recursion and longest-increasing-sequence code. An earlier struct-method version was slower for this same layout reason, and anchoring has its own variant of that code.

**The `-b` row** does not compare like with like. The old API had no whitespace-aware renderer with labels, so the baseline is `line_diff_unified_hunk_body` (hunks only, no `---` / `+++` lines, no function names). The new row adds labels and the function-name pass on top of the same, unchanged body engine.

For context, `git diff --no-index` on the same Rust files takes 3.0–3.5 ms per run on this VM across these options, while `git --version` alone takes 0.7 ms. That includes process start, reading both files and hashing them for the `index` line, so it does not compare directly with the in-process rows.

```bash
cargo bench -p grit-lib --bench diff
cargo bench -p grit-lib --bench userdiff_hunk_header
cargo bench -p grit-lib --bench merge_content
```
